1 /*- 2 * Copyright (c) 2016-2019 3 * Netflix Inc. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_ipsec.h" 34 #include "opt_tcpdebug.h" 35 36 #include <sys/param.h> 37 #include <sys/module.h> 38 #include <sys/kernel.h> 39 #ifdef TCP_HHOOK 40 #include <sys/hhook.h> 41 #endif 42 #include <sys/lock.h> 43 #include <sys/malloc.h> 44 #include <sys/lock.h> 45 #include <sys/mutex.h> 46 #include <sys/mbuf.h> 47 #include <sys/proc.h> /* for proc0 declaration */ 48 #ifdef NETFLIX_STATS 49 #include <sys/qmath.h> 50 #endif 51 #include <sys/socket.h> 52 #include <sys/socketvar.h> 53 #include <sys/sysctl.h> 54 #include <sys/systm.h> 55 #include <sys/tree.h> 56 #ifdef NETFLIX_STATS 57 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 58 #endif 59 #include <sys/refcount.h> 60 #include <sys/queue.h> 61 #include <sys/smp.h> 62 #include <sys/kthread.h> 63 #include <sys/kern_prefetch.h> 64 65 #include <vm/uma.h> 66 67 #include <net/route.h> 68 #include <net/vnet.h> 69 70 #define TCPSTATES /* for logging */ 71 72 #include <netinet/in.h> 73 #include <netinet/in_kdtrace.h> 74 #include <netinet/in_pcb.h> 75 #include <netinet/ip.h> 76 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 77 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 78 #include <netinet/ip_var.h> 79 #include <netinet/ip6.h> 80 #include <netinet6/in6_pcb.h> 81 #include <netinet6/ip6_var.h> 82 #define TCPOUTFLAGS 83 #include <netinet/tcp.h> 84 #include <netinet/tcp_fsm.h> 85 #include <netinet/tcp_log_buf.h> 86 #include <netinet/tcp_seq.h> 87 #include <netinet/tcp_timer.h> 88 #include <netinet/tcp_var.h> 89 #include <netinet/tcp_hpts.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/cc/cc.h> 92 #include <netinet/tcp_fastopen.h> 93 #ifdef TCPDEBUG 94 #include <netinet/tcp_debug.h> 95 #endif /* TCPDEBUG */ 96 #ifdef TCP_OFFLOAD 97 #include <netinet/tcp_offload.h> 98 #endif 99 #ifdef INET6 100 #include <netinet6/tcp6_var.h> 101 #endif 102 103 #include <netipsec/ipsec_support.h> 104 105 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 106 #include <netipsec/ipsec.h> 107 #include <netipsec/ipsec6.h> 108 #endif /* IPSEC */ 109 110 #include <netinet/udp.h> 111 #include <netinet/udp_var.h> 112 #include <machine/in_cksum.h> 113 114 #ifdef MAC 115 #include <security/mac/mac_framework.h> 116 #endif 117 #include "sack_filter.h" 118 #include "tcp_rack.h" 119 #include "rack_bbr_common.h" 120 121 uma_zone_t rack_zone; 122 uma_zone_t rack_pcb_zone; 123 124 #ifndef TICKS2SBT 125 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) 126 #endif 127 128 struct sysctl_ctx_list rack_sysctl_ctx; 129 struct sysctl_oid *rack_sysctl_root; 130 131 #ifndef TCPHPTS 132 #error "fatal error missing option TCPHSTS in the build" 133 #endif 134 135 #define CUM_ACKED 1 136 #define SACKED 2 137 138 /* 139 * The RACK module incorporates a number of 140 * TCP ideas that have been put out into the IETF 141 * over the last few years: 142 * - Matt Mathis's Rate Halving which slowly drops 143 * the congestion window so that the ack clock can 144 * be maintained during a recovery. 145 * - Yuchung Cheng's RACK TCP (for which its named) that 146 * will stop us using the number of dup acks and instead 147 * use time as the gage of when we retransmit. 148 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 149 * of Dukkipati et.al. 150 * RACK depends on SACK, so if an endpoint arrives that 151 * cannot do SACK the state machine below will shuttle the 152 * connection back to using the "default" TCP stack that is 153 * in FreeBSD. 154 * 155 * To implement RACK the original TCP stack was first decomposed 156 * into a functional state machine with individual states 157 * for each of the possible TCP connection states. The do_segement 158 * functions role in life is to mandate the connection supports SACK 159 * initially and then assure that the RACK state matches the conenction 160 * state before calling the states do_segment function. Each 161 * state is simplified due to the fact that the original do_segment 162 * has been decomposed and we *know* what state we are in (no 163 * switches on the state) and all tests for SACK are gone. This 164 * greatly simplifies what each state does. 165 * 166 * TCP output is also over-written with a new version since it 167 * must maintain the new rack scoreboard. 168 * 169 */ 170 static int32_t rack_precache = 1; 171 static int32_t rack_tlp_thresh = 1; 172 static int32_t rack_reorder_thresh = 2; 173 static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 174 * - 60 seconds */ 175 static int32_t rack_pkt_delay = 1; 176 static int32_t rack_inc_var = 0;/* For TLP */ 177 static int32_t rack_reduce_largest_on_idle = 0; 178 static int32_t rack_min_pace_time = 0; 179 static int32_t rack_min_pace_time_seg_req=6; 180 static int32_t rack_early_recovery = 1; 181 static int32_t rack_early_recovery_max_seg = 6; 182 static int32_t rack_send_a_lot_in_prr = 1; 183 static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ 184 static int32_t rack_tlp_in_recovery = 1; /* Can we do TLP in recovery? */ 185 static int32_t rack_verbose_logging = 0; 186 static int32_t rack_ignore_data_after_close = 1; 187 static int32_t rack_map_entries_limit = 1024; 188 static int32_t rack_map_split_limit = 256; 189 190 /* 191 * Currently regular tcp has a rto_min of 30ms 192 * the backoff goes 12 times so that ends up 193 * being a total of 122.850 seconds before a 194 * connection is killed. 195 */ 196 static int32_t rack_tlp_min = 10; 197 static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ 198 static int32_t rack_rto_max = 30000; /* 30 seconds */ 199 static const int32_t rack_free_cache = 2; 200 static int32_t rack_hptsi_segments = 40; 201 static int32_t rack_rate_sample_method = USE_RTT_LOW; 202 static int32_t rack_pace_every_seg = 1; 203 static int32_t rack_delayed_ack_time = 200; /* 200ms */ 204 static int32_t rack_slot_reduction = 4; 205 static int32_t rack_lower_cwnd_at_tlp = 0; 206 static int32_t rack_use_proportional_reduce = 0; 207 static int32_t rack_proportional_rate = 10; 208 static int32_t rack_tlp_max_resend = 2; 209 static int32_t rack_limited_retran = 0; 210 static int32_t rack_always_send_oldest = 0; 211 static int32_t rack_sack_block_limit = 128; 212 static int32_t rack_use_sack_filter = 1; 213 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; 214 215 /* Rack specific counters */ 216 counter_u64_t rack_badfr; 217 counter_u64_t rack_badfr_bytes; 218 counter_u64_t rack_rtm_prr_retran; 219 counter_u64_t rack_rtm_prr_newdata; 220 counter_u64_t rack_timestamp_mismatch; 221 counter_u64_t rack_reorder_seen; 222 counter_u64_t rack_paced_segments; 223 counter_u64_t rack_unpaced_segments; 224 counter_u64_t rack_saw_enobuf; 225 counter_u64_t rack_saw_enetunreach; 226 227 /* Tail loss probe counters */ 228 counter_u64_t rack_tlp_tot; 229 counter_u64_t rack_tlp_newdata; 230 counter_u64_t rack_tlp_retran; 231 counter_u64_t rack_tlp_retran_bytes; 232 counter_u64_t rack_tlp_retran_fail; 233 counter_u64_t rack_to_tot; 234 counter_u64_t rack_to_arm_rack; 235 counter_u64_t rack_to_arm_tlp; 236 counter_u64_t rack_to_alloc; 237 counter_u64_t rack_to_alloc_hard; 238 counter_u64_t rack_to_alloc_emerg; 239 counter_u64_t rack_to_alloc_limited; 240 counter_u64_t rack_alloc_limited_conns; 241 counter_u64_t rack_split_limited; 242 243 counter_u64_t rack_sack_proc_all; 244 counter_u64_t rack_sack_proc_short; 245 counter_u64_t rack_sack_proc_restart; 246 counter_u64_t rack_runt_sacks; 247 counter_u64_t rack_used_tlpmethod; 248 counter_u64_t rack_used_tlpmethod2; 249 counter_u64_t rack_enter_tlp_calc; 250 counter_u64_t rack_input_idle_reduces; 251 counter_u64_t rack_tlp_does_nada; 252 253 /* Temp CPU counters */ 254 counter_u64_t rack_find_high; 255 256 counter_u64_t rack_progress_drops; 257 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; 258 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 259 260 /* 261 * This was originally defined in tcp_timer.c, but is now reproduced here given 262 * the unification of the SYN and non-SYN retransmit timer exponents combined 263 * with wanting to retain previous behaviour for previously deployed stack 264 * versions. 265 */ 266 int tcp_syn_backoff[TCP_MAXRXTSHIFT + 1] = 267 { 1, 1, 1, 1, 1, 2, 4, 8, 16, 32, 64, 64, 64 }; 268 269 static void 270 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); 271 272 static int 273 rack_process_ack(struct mbuf *m, struct tcphdr *th, 274 struct socket *so, struct tcpcb *tp, struct tcpopt *to, 275 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); 276 static int 277 rack_process_data(struct mbuf *m, struct tcphdr *th, 278 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 279 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 280 static void 281 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, 282 struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); 283 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); 284 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, 285 uint8_t limit_type); 286 static struct rack_sendmap * 287 rack_check_recovery_mode(struct tcpcb *tp, 288 uint32_t tsused); 289 static void 290 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, 291 uint32_t type); 292 static void rack_counter_destroy(void); 293 static int 294 rack_ctloutput(struct socket *so, struct sockopt *sopt, 295 struct inpcb *inp, struct tcpcb *tp); 296 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); 297 static void 298 rack_do_segment(struct mbuf *m, struct tcphdr *th, 299 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 300 uint8_t iptos); 301 static void rack_dtor(void *mem, int32_t size, void *arg); 302 static void 303 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 304 uint32_t t, uint32_t cts); 305 static struct rack_sendmap * 306 rack_find_high_nonack(struct tcp_rack *rack, 307 struct rack_sendmap *rsm); 308 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); 309 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); 310 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); 311 static int 312 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 313 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 314 static int32_t rack_handoff_ok(struct tcpcb *tp); 315 static int32_t rack_init(struct tcpcb *tp); 316 static void rack_init_sysctls(void); 317 static void 318 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, 319 struct tcphdr *th); 320 static void 321 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 322 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 323 uint8_t pass, struct rack_sendmap *hintrsm); 324 static void 325 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, 326 struct rack_sendmap *rsm); 327 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num); 328 static int32_t rack_output(struct tcpcb *tp); 329 static void 330 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, 331 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 332 uint8_t iptos, int32_t nxt_pkt, struct timeval *tv); 333 334 static uint32_t 335 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, 336 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, 337 uint32_t cts); 338 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); 339 static void rack_remxt_tmr(struct tcpcb *tp); 340 static int 341 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 342 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 343 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); 344 static int32_t rack_stopall(struct tcpcb *tp); 345 static void 346 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, 347 uint32_t delta); 348 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); 349 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); 350 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); 351 static uint32_t 352 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 353 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); 354 static void 355 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 356 struct rack_sendmap *rsm, uint32_t ts); 357 static int 358 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 359 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type); 360 static int32_t tcp_addrack(module_t mod, int32_t type, void *data); 361 static void 362 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, 363 struct tcpcb *tp, int32_t * ret_val); 364 static int 365 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, 366 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 367 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 368 static int 369 rack_do_closing(struct mbuf *m, struct tcphdr *th, 370 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 371 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 372 static void rack_do_drop(struct mbuf *m, struct tcpcb *tp); 373 static void 374 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, 375 struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val); 376 static void 377 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, 378 struct tcphdr *th, int32_t rstreason, int32_t tlen); 379 static int 380 rack_do_established(struct mbuf *m, struct tcphdr *th, 381 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 382 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 383 static int 384 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, 385 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 386 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt); 387 static int 388 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, 389 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 390 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 391 static int 392 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, 393 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 394 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 395 static int 396 rack_do_lastack(struct mbuf *m, struct tcphdr *th, 397 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 398 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 399 static int 400 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, 401 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 402 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 403 static int 404 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, 405 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, 406 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); 407 static int 408 rack_drop_checks(struct tcpopt *to, struct mbuf *m, 409 struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, 410 int32_t * drop_hdrlen, int32_t * ret_val); 411 static int 412 rack_process_rst(struct mbuf *m, struct tcphdr *th, 413 struct socket *so, struct tcpcb *tp); 414 struct rack_sendmap * 415 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, 416 uint32_t tsused); 417 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt); 418 static void 419 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); 420 421 static int 422 rack_ts_check(struct mbuf *m, struct tcphdr *th, 423 struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val); 424 425 int32_t rack_clear_counter=0; 426 427 428 static int 429 sysctl_rack_clear(SYSCTL_HANDLER_ARGS) 430 { 431 uint32_t stat; 432 int32_t error; 433 434 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); 435 if (error || req->newptr == NULL) 436 return error; 437 438 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 439 if (error) 440 return (error); 441 if (stat == 1) { 442 #ifdef INVARIANTS 443 printf("Clearing RACK counters\n"); 444 #endif 445 counter_u64_zero(rack_badfr); 446 counter_u64_zero(rack_badfr_bytes); 447 counter_u64_zero(rack_rtm_prr_retran); 448 counter_u64_zero(rack_rtm_prr_newdata); 449 counter_u64_zero(rack_timestamp_mismatch); 450 counter_u64_zero(rack_reorder_seen); 451 counter_u64_zero(rack_tlp_tot); 452 counter_u64_zero(rack_tlp_newdata); 453 counter_u64_zero(rack_tlp_retran); 454 counter_u64_zero(rack_tlp_retran_bytes); 455 counter_u64_zero(rack_tlp_retran_fail); 456 counter_u64_zero(rack_to_tot); 457 counter_u64_zero(rack_to_arm_rack); 458 counter_u64_zero(rack_to_arm_tlp); 459 counter_u64_zero(rack_paced_segments); 460 counter_u64_zero(rack_unpaced_segments); 461 counter_u64_zero(rack_saw_enobuf); 462 counter_u64_zero(rack_saw_enetunreach); 463 counter_u64_zero(rack_to_alloc_hard); 464 counter_u64_zero(rack_to_alloc_emerg); 465 counter_u64_zero(rack_sack_proc_all); 466 counter_u64_zero(rack_sack_proc_short); 467 counter_u64_zero(rack_sack_proc_restart); 468 counter_u64_zero(rack_to_alloc); 469 counter_u64_zero(rack_to_alloc_limited); 470 counter_u64_zero(rack_alloc_limited_conns); 471 counter_u64_zero(rack_split_limited); 472 counter_u64_zero(rack_find_high); 473 counter_u64_zero(rack_runt_sacks); 474 counter_u64_zero(rack_used_tlpmethod); 475 counter_u64_zero(rack_used_tlpmethod2); 476 counter_u64_zero(rack_enter_tlp_calc); 477 counter_u64_zero(rack_progress_drops); 478 counter_u64_zero(rack_tlp_does_nada); 479 } 480 rack_clear_counter = 0; 481 return (0); 482 } 483 484 485 486 static void 487 rack_init_sysctls() 488 { 489 SYSCTL_ADD_S32(&rack_sysctl_ctx, 490 SYSCTL_CHILDREN(rack_sysctl_root), 491 OID_AUTO, "map_limit", CTLFLAG_RW, 492 &rack_map_entries_limit , 1024, 493 "Is there a limit on how big the sendmap can grow? "); 494 495 SYSCTL_ADD_S32(&rack_sysctl_ctx, 496 SYSCTL_CHILDREN(rack_sysctl_root), 497 OID_AUTO, "map_splitlimit", CTLFLAG_RW, 498 &rack_map_split_limit , 256, 499 "Is there a limit on how much splitting a peer can do?"); 500 501 SYSCTL_ADD_S32(&rack_sysctl_ctx, 502 SYSCTL_CHILDREN(rack_sysctl_root), 503 OID_AUTO, "rate_sample_method", CTLFLAG_RW, 504 &rack_rate_sample_method , USE_RTT_LOW, 505 "What method should we use for rate sampling 0=high, 1=low "); 506 SYSCTL_ADD_S32(&rack_sysctl_ctx, 507 SYSCTL_CHILDREN(rack_sysctl_root), 508 OID_AUTO, "data_after_close", CTLFLAG_RW, 509 &rack_ignore_data_after_close, 0, 510 "Do we hold off sending a RST until all pending data is ack'd"); 511 SYSCTL_ADD_S32(&rack_sysctl_ctx, 512 SYSCTL_CHILDREN(rack_sysctl_root), 513 OID_AUTO, "tlpmethod", CTLFLAG_RW, 514 &rack_tlp_threshold_use, TLP_USE_TWO_ONE, 515 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); 516 SYSCTL_ADD_S32(&rack_sysctl_ctx, 517 SYSCTL_CHILDREN(rack_sysctl_root), 518 OID_AUTO, "min_pace_time", CTLFLAG_RW, 519 &rack_min_pace_time, 0, 520 "Should we enforce a minimum pace time of 1ms"); 521 SYSCTL_ADD_S32(&rack_sysctl_ctx, 522 SYSCTL_CHILDREN(rack_sysctl_root), 523 OID_AUTO, "min_pace_segs", CTLFLAG_RW, 524 &rack_min_pace_time_seg_req, 6, 525 "How many segments have to be in the len to enforce min-pace-time"); 526 SYSCTL_ADD_S32(&rack_sysctl_ctx, 527 SYSCTL_CHILDREN(rack_sysctl_root), 528 OID_AUTO, "idle_reduce_high", CTLFLAG_RW, 529 &rack_reduce_largest_on_idle, 0, 530 "Should we reduce the largest cwnd seen to IW on idle reduction"); 531 SYSCTL_ADD_S32(&rack_sysctl_ctx, 532 SYSCTL_CHILDREN(rack_sysctl_root), 533 OID_AUTO, "bb_verbose", CTLFLAG_RW, 534 &rack_verbose_logging, 0, 535 "Should RACK black box logging be verbose"); 536 SYSCTL_ADD_S32(&rack_sysctl_ctx, 537 SYSCTL_CHILDREN(rack_sysctl_root), 538 OID_AUTO, "sackfiltering", CTLFLAG_RW, 539 &rack_use_sack_filter, 1, 540 "Do we use sack filtering?"); 541 SYSCTL_ADD_S32(&rack_sysctl_ctx, 542 SYSCTL_CHILDREN(rack_sysctl_root), 543 OID_AUTO, "delayed_ack", CTLFLAG_RW, 544 &rack_delayed_ack_time, 200, 545 "Delayed ack time (200ms)"); 546 SYSCTL_ADD_S32(&rack_sysctl_ctx, 547 SYSCTL_CHILDREN(rack_sysctl_root), 548 OID_AUTO, "tlpminto", CTLFLAG_RW, 549 &rack_tlp_min, 10, 550 "TLP minimum timeout per the specification (10ms)"); 551 SYSCTL_ADD_S32(&rack_sysctl_ctx, 552 SYSCTL_CHILDREN(rack_sysctl_root), 553 OID_AUTO, "precache", CTLFLAG_RW, 554 &rack_precache, 0, 555 "Where should we precache the mcopy (0 is not at all)"); 556 SYSCTL_ADD_S32(&rack_sysctl_ctx, 557 SYSCTL_CHILDREN(rack_sysctl_root), 558 OID_AUTO, "sblklimit", CTLFLAG_RW, 559 &rack_sack_block_limit, 128, 560 "When do we start paying attention to small sack blocks"); 561 SYSCTL_ADD_S32(&rack_sysctl_ctx, 562 SYSCTL_CHILDREN(rack_sysctl_root), 563 OID_AUTO, "send_oldest", CTLFLAG_RW, 564 &rack_always_send_oldest, 1, 565 "Should we always send the oldest TLP and RACK-TLP"); 566 SYSCTL_ADD_S32(&rack_sysctl_ctx, 567 SYSCTL_CHILDREN(rack_sysctl_root), 568 OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW, 569 &rack_tlp_in_recovery, 1, 570 "Can we do a TLP during recovery?"); 571 SYSCTL_ADD_S32(&rack_sysctl_ctx, 572 SYSCTL_CHILDREN(rack_sysctl_root), 573 OID_AUTO, "rack_tlimit", CTLFLAG_RW, 574 &rack_limited_retran, 0, 575 "How many times can a rack timeout drive out sends"); 576 SYSCTL_ADD_S32(&rack_sysctl_ctx, 577 SYSCTL_CHILDREN(rack_sysctl_root), 578 OID_AUTO, "minrto", CTLFLAG_RW, 579 &rack_rto_min, 0, 580 "Minimum RTO in ms -- set with caution below 1000 due to TLP"); 581 SYSCTL_ADD_S32(&rack_sysctl_ctx, 582 SYSCTL_CHILDREN(rack_sysctl_root), 583 OID_AUTO, "maxrto", CTLFLAG_RW, 584 &rack_rto_max, 0, 585 "Maxiumum RTO in ms -- should be at least as large as min_rto"); 586 SYSCTL_ADD_S32(&rack_sysctl_ctx, 587 SYSCTL_CHILDREN(rack_sysctl_root), 588 OID_AUTO, "tlp_retry", CTLFLAG_RW, 589 &rack_tlp_max_resend, 2, 590 "How many times does TLP retry a single segment or multiple with no ACK"); 591 SYSCTL_ADD_S32(&rack_sysctl_ctx, 592 SYSCTL_CHILDREN(rack_sysctl_root), 593 OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, 594 &rack_use_proportional_reduce, 0, 595 "Should we proportionaly reduce cwnd based on the number of losses "); 596 SYSCTL_ADD_S32(&rack_sysctl_ctx, 597 SYSCTL_CHILDREN(rack_sysctl_root), 598 OID_AUTO, "recovery_prop", CTLFLAG_RW, 599 &rack_proportional_rate, 10, 600 "What percent reduction per loss"); 601 SYSCTL_ADD_S32(&rack_sysctl_ctx, 602 SYSCTL_CHILDREN(rack_sysctl_root), 603 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, 604 &rack_lower_cwnd_at_tlp, 0, 605 "When a TLP completes a retran should we enter recovery?"); 606 SYSCTL_ADD_S32(&rack_sysctl_ctx, 607 SYSCTL_CHILDREN(rack_sysctl_root), 608 OID_AUTO, "hptsi_reduces", CTLFLAG_RW, 609 &rack_slot_reduction, 4, 610 "When setting a slot should we reduce by divisor"); 611 SYSCTL_ADD_S32(&rack_sysctl_ctx, 612 SYSCTL_CHILDREN(rack_sysctl_root), 613 OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, 614 &rack_pace_every_seg, 1, 615 "Should we pace out every segment hptsi"); 616 SYSCTL_ADD_S32(&rack_sysctl_ctx, 617 SYSCTL_CHILDREN(rack_sysctl_root), 618 OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, 619 &rack_hptsi_segments, 6, 620 "Should we pace out only a limited size of segments"); 621 SYSCTL_ADD_S32(&rack_sysctl_ctx, 622 SYSCTL_CHILDREN(rack_sysctl_root), 623 OID_AUTO, "prr_sendalot", CTLFLAG_RW, 624 &rack_send_a_lot_in_prr, 1, 625 "Send a lot in prr"); 626 SYSCTL_ADD_S32(&rack_sysctl_ctx, 627 SYSCTL_CHILDREN(rack_sysctl_root), 628 OID_AUTO, "minto", CTLFLAG_RW, 629 &rack_min_to, 1, 630 "Minimum rack timeout in milliseconds"); 631 SYSCTL_ADD_S32(&rack_sysctl_ctx, 632 SYSCTL_CHILDREN(rack_sysctl_root), 633 OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW, 634 &rack_early_recovery_max_seg, 6, 635 "Max segments in early recovery"); 636 SYSCTL_ADD_S32(&rack_sysctl_ctx, 637 SYSCTL_CHILDREN(rack_sysctl_root), 638 OID_AUTO, "earlyrecovery", CTLFLAG_RW, 639 &rack_early_recovery, 1, 640 "Do we do early recovery with rack"); 641 SYSCTL_ADD_S32(&rack_sysctl_ctx, 642 SYSCTL_CHILDREN(rack_sysctl_root), 643 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 644 &rack_reorder_thresh, 2, 645 "What factor for rack will be added when seeing reordering (shift right)"); 646 SYSCTL_ADD_S32(&rack_sysctl_ctx, 647 SYSCTL_CHILDREN(rack_sysctl_root), 648 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 649 &rack_tlp_thresh, 1, 650 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 651 SYSCTL_ADD_S32(&rack_sysctl_ctx, 652 SYSCTL_CHILDREN(rack_sysctl_root), 653 OID_AUTO, "reorder_fade", CTLFLAG_RW, 654 &rack_reorder_fade, 0, 655 "Does reorder detection fade, if so how many ms (0 means never)"); 656 SYSCTL_ADD_S32(&rack_sysctl_ctx, 657 SYSCTL_CHILDREN(rack_sysctl_root), 658 OID_AUTO, "pktdelay", CTLFLAG_RW, 659 &rack_pkt_delay, 1, 660 "Extra RACK time (in ms) besides reordering thresh"); 661 SYSCTL_ADD_S32(&rack_sysctl_ctx, 662 SYSCTL_CHILDREN(rack_sysctl_root), 663 OID_AUTO, "inc_var", CTLFLAG_RW, 664 &rack_inc_var, 0, 665 "Should rack add to the TLP timer the variance in rtt calculation"); 666 rack_badfr = counter_u64_alloc(M_WAITOK); 667 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 668 SYSCTL_CHILDREN(rack_sysctl_root), 669 OID_AUTO, "badfr", CTLFLAG_RD, 670 &rack_badfr, "Total number of bad FRs"); 671 rack_badfr_bytes = counter_u64_alloc(M_WAITOK); 672 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 673 SYSCTL_CHILDREN(rack_sysctl_root), 674 OID_AUTO, "badfr_bytes", CTLFLAG_RD, 675 &rack_badfr_bytes, "Total number of bad FRs"); 676 rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); 677 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 678 SYSCTL_CHILDREN(rack_sysctl_root), 679 OID_AUTO, "prrsndret", CTLFLAG_RD, 680 &rack_rtm_prr_retran, 681 "Total number of prr based retransmits"); 682 rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); 683 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 684 SYSCTL_CHILDREN(rack_sysctl_root), 685 OID_AUTO, "prrsndnew", CTLFLAG_RD, 686 &rack_rtm_prr_newdata, 687 "Total number of prr based new transmits"); 688 rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); 689 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 690 SYSCTL_CHILDREN(rack_sysctl_root), 691 OID_AUTO, "tsnf", CTLFLAG_RD, 692 &rack_timestamp_mismatch, 693 "Total number of timestamps that we could not find the reported ts"); 694 rack_find_high = counter_u64_alloc(M_WAITOK); 695 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 696 SYSCTL_CHILDREN(rack_sysctl_root), 697 OID_AUTO, "findhigh", CTLFLAG_RD, 698 &rack_find_high, 699 "Total number of FIN causing find-high"); 700 rack_reorder_seen = counter_u64_alloc(M_WAITOK); 701 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 702 SYSCTL_CHILDREN(rack_sysctl_root), 703 OID_AUTO, "reordering", CTLFLAG_RD, 704 &rack_reorder_seen, 705 "Total number of times we added delay due to reordering"); 706 rack_tlp_tot = counter_u64_alloc(M_WAITOK); 707 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 708 SYSCTL_CHILDREN(rack_sysctl_root), 709 OID_AUTO, "tlp_to_total", CTLFLAG_RD, 710 &rack_tlp_tot, 711 "Total number of tail loss probe expirations"); 712 rack_tlp_newdata = counter_u64_alloc(M_WAITOK); 713 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 714 SYSCTL_CHILDREN(rack_sysctl_root), 715 OID_AUTO, "tlp_new", CTLFLAG_RD, 716 &rack_tlp_newdata, 717 "Total number of tail loss probe sending new data"); 718 719 rack_tlp_retran = counter_u64_alloc(M_WAITOK); 720 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 721 SYSCTL_CHILDREN(rack_sysctl_root), 722 OID_AUTO, "tlp_retran", CTLFLAG_RD, 723 &rack_tlp_retran, 724 "Total number of tail loss probe sending retransmitted data"); 725 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); 726 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 727 SYSCTL_CHILDREN(rack_sysctl_root), 728 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, 729 &rack_tlp_retran_bytes, 730 "Total bytes of tail loss probe sending retransmitted data"); 731 rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); 732 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 733 SYSCTL_CHILDREN(rack_sysctl_root), 734 OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, 735 &rack_tlp_retran_fail, 736 "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); 737 rack_to_tot = counter_u64_alloc(M_WAITOK); 738 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 739 SYSCTL_CHILDREN(rack_sysctl_root), 740 OID_AUTO, "rack_to_tot", CTLFLAG_RD, 741 &rack_to_tot, 742 "Total number of times the rack to expired?"); 743 rack_to_arm_rack = counter_u64_alloc(M_WAITOK); 744 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 745 SYSCTL_CHILDREN(rack_sysctl_root), 746 OID_AUTO, "arm_rack", CTLFLAG_RD, 747 &rack_to_arm_rack, 748 "Total number of times the rack timer armed?"); 749 rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); 750 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 751 SYSCTL_CHILDREN(rack_sysctl_root), 752 OID_AUTO, "arm_tlp", CTLFLAG_RD, 753 &rack_to_arm_tlp, 754 "Total number of times the tlp timer armed?"); 755 rack_paced_segments = counter_u64_alloc(M_WAITOK); 756 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 757 SYSCTL_CHILDREN(rack_sysctl_root), 758 OID_AUTO, "paced", CTLFLAG_RD, 759 &rack_paced_segments, 760 "Total number of times a segment send caused hptsi"); 761 rack_unpaced_segments = counter_u64_alloc(M_WAITOK); 762 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 763 SYSCTL_CHILDREN(rack_sysctl_root), 764 OID_AUTO, "unpaced", CTLFLAG_RD, 765 &rack_unpaced_segments, 766 "Total number of times a segment did not cause hptsi"); 767 rack_saw_enobuf = counter_u64_alloc(M_WAITOK); 768 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 769 SYSCTL_CHILDREN(rack_sysctl_root), 770 OID_AUTO, "saw_enobufs", CTLFLAG_RD, 771 &rack_saw_enobuf, 772 "Total number of times a segment did not cause hptsi"); 773 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); 774 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 775 SYSCTL_CHILDREN(rack_sysctl_root), 776 OID_AUTO, "saw_enetunreach", CTLFLAG_RD, 777 &rack_saw_enetunreach, 778 "Total number of times a segment did not cause hptsi"); 779 rack_to_alloc = counter_u64_alloc(M_WAITOK); 780 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 781 SYSCTL_CHILDREN(rack_sysctl_root), 782 OID_AUTO, "allocs", CTLFLAG_RD, 783 &rack_to_alloc, 784 "Total allocations of tracking structures"); 785 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); 786 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 787 SYSCTL_CHILDREN(rack_sysctl_root), 788 OID_AUTO, "allochard", CTLFLAG_RD, 789 &rack_to_alloc_hard, 790 "Total allocations done with sleeping the hard way"); 791 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); 792 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 793 SYSCTL_CHILDREN(rack_sysctl_root), 794 OID_AUTO, "allocemerg", CTLFLAG_RD, 795 &rack_to_alloc_emerg, 796 "Total allocations done from emergency cache"); 797 rack_to_alloc_limited = counter_u64_alloc(M_WAITOK); 798 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 799 SYSCTL_CHILDREN(rack_sysctl_root), 800 OID_AUTO, "alloc_limited", CTLFLAG_RD, 801 &rack_to_alloc_limited, 802 "Total allocations dropped due to limit"); 803 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); 804 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 805 SYSCTL_CHILDREN(rack_sysctl_root), 806 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, 807 &rack_alloc_limited_conns, 808 "Connections with allocations dropped due to limit"); 809 rack_split_limited = counter_u64_alloc(M_WAITOK); 810 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 811 SYSCTL_CHILDREN(rack_sysctl_root), 812 OID_AUTO, "split_limited", CTLFLAG_RD, 813 &rack_split_limited, 814 "Split allocations dropped due to limit"); 815 rack_sack_proc_all = counter_u64_alloc(M_WAITOK); 816 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 817 SYSCTL_CHILDREN(rack_sysctl_root), 818 OID_AUTO, "sack_long", CTLFLAG_RD, 819 &rack_sack_proc_all, 820 "Total times we had to walk whole list for sack processing"); 821 822 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); 823 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 824 SYSCTL_CHILDREN(rack_sysctl_root), 825 OID_AUTO, "sack_restart", CTLFLAG_RD, 826 &rack_sack_proc_restart, 827 "Total times we had to walk whole list due to a restart"); 828 rack_sack_proc_short = counter_u64_alloc(M_WAITOK); 829 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 830 SYSCTL_CHILDREN(rack_sysctl_root), 831 OID_AUTO, "sack_short", CTLFLAG_RD, 832 &rack_sack_proc_short, 833 "Total times we took shortcut for sack processing"); 834 rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); 835 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 836 SYSCTL_CHILDREN(rack_sysctl_root), 837 OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, 838 &rack_enter_tlp_calc, 839 "Total times we called calc-tlp"); 840 rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); 841 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 842 SYSCTL_CHILDREN(rack_sysctl_root), 843 OID_AUTO, "hit_tlp_method", CTLFLAG_RD, 844 &rack_used_tlpmethod, 845 "Total number of runt sacks"); 846 rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); 847 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 848 SYSCTL_CHILDREN(rack_sysctl_root), 849 OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, 850 &rack_used_tlpmethod2, 851 "Total number of runt sacks 2"); 852 rack_runt_sacks = counter_u64_alloc(M_WAITOK); 853 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 854 SYSCTL_CHILDREN(rack_sysctl_root), 855 OID_AUTO, "runtsacks", CTLFLAG_RD, 856 &rack_runt_sacks, 857 "Total number of runt sacks"); 858 rack_progress_drops = counter_u64_alloc(M_WAITOK); 859 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 860 SYSCTL_CHILDREN(rack_sysctl_root), 861 OID_AUTO, "prog_drops", CTLFLAG_RD, 862 &rack_progress_drops, 863 "Total number of progress drops"); 864 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); 865 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 866 SYSCTL_CHILDREN(rack_sysctl_root), 867 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, 868 &rack_input_idle_reduces, 869 "Total number of idle reductions on input"); 870 rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); 871 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, 872 SYSCTL_CHILDREN(rack_sysctl_root), 873 OID_AUTO, "tlp_nada", CTLFLAG_RD, 874 &rack_tlp_does_nada, 875 "Total number of nada tlp calls"); 876 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 877 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 878 OID_AUTO, "outsize", CTLFLAG_RD, 879 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); 880 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); 881 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 882 OID_AUTO, "opts", CTLFLAG_RD, 883 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); 884 SYSCTL_ADD_PROC(&rack_sysctl_ctx, 885 SYSCTL_CHILDREN(rack_sysctl_root), 886 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 887 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); 888 } 889 890 static inline int32_t 891 rack_progress_timeout_check(struct tcpcb *tp) 892 { 893 #ifdef NETFLIX_PROGRESS 894 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 895 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 896 /* 897 * There is an assumption that the caller 898 * will drop the connection so we will 899 * increment the counters here. 900 */ 901 struct tcp_rack *rack; 902 rack = (struct tcp_rack *)tp->t_fb_ptr; 903 counter_u64_add(rack_progress_drops, 1); 904 TCPSTAT_INC(tcps_progdrops); 905 rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); 906 return (1); 907 } 908 } 909 #endif 910 return (0); 911 } 912 913 914 static void 915 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 916 { 917 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 918 union tcp_log_stackspecific log; 919 920 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 921 log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); 922 log.u_bbr.flex2 = to; 923 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 924 log.u_bbr.flex4 = slot; 925 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; 926 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 927 log.u_bbr.flex8 = which; 928 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 929 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 930 TCP_LOG_EVENT(rack->rc_tp, NULL, 931 &rack->rc_inp->inp_socket->so_rcv, 932 &rack->rc_inp->inp_socket->so_snd, 933 BBR_LOG_TIMERSTAR, 0, 934 0, &log, false); 935 } 936 } 937 938 static void 939 rack_log_to_event(struct tcp_rack *rack, int32_t to_num) 940 { 941 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 942 union tcp_log_stackspecific log; 943 944 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 945 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 946 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 947 log.u_bbr.flex8 = to_num; 948 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; 949 log.u_bbr.flex2 = rack->rc_rack_rtt; 950 TCP_LOG_EVENT(rack->rc_tp, NULL, 951 &rack->rc_inp->inp_socket->so_rcv, 952 &rack->rc_inp->inp_socket->so_snd, 953 BBR_LOG_RTO, 0, 954 0, &log, false); 955 } 956 } 957 958 static void 959 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, 960 uint32_t o_srtt, uint32_t o_var) 961 { 962 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 963 union tcp_log_stackspecific log; 964 965 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 966 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 967 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 968 log.u_bbr.flex1 = t; 969 log.u_bbr.flex2 = o_srtt; 970 log.u_bbr.flex3 = o_var; 971 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; 972 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; 973 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; 974 log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; 975 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; 976 TCP_LOG_EVENT(tp, NULL, 977 &rack->rc_inp->inp_socket->so_rcv, 978 &rack->rc_inp->inp_socket->so_snd, 979 BBR_LOG_BBRRTT, 0, 980 0, &log, false); 981 } 982 } 983 984 static void 985 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) 986 { 987 /* 988 * Log the rtt sample we are 989 * applying to the srtt algorithm in 990 * useconds. 991 */ 992 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 993 union tcp_log_stackspecific log; 994 struct timeval tv; 995 996 memset(&log, 0, sizeof(log)); 997 /* Convert our ms to a microsecond */ 998 log.u_bbr.flex1 = rtt * 1000; 999 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 1000 TCP_LOG_EVENTP(rack->rc_tp, NULL, 1001 &rack->rc_inp->inp_socket->so_rcv, 1002 &rack->rc_inp->inp_socket->so_snd, 1003 TCP_LOG_RTT, 0, 1004 0, &log, false, &tv); 1005 } 1006 } 1007 1008 1009 static inline void 1010 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) 1011 { 1012 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { 1013 union tcp_log_stackspecific log; 1014 1015 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1016 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1017 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1018 log.u_bbr.flex1 = line; 1019 log.u_bbr.flex2 = tick; 1020 log.u_bbr.flex3 = tp->t_maxunacktime; 1021 log.u_bbr.flex4 = tp->t_acktime; 1022 log.u_bbr.flex8 = event; 1023 TCP_LOG_EVENT(tp, NULL, 1024 &rack->rc_inp->inp_socket->so_rcv, 1025 &rack->rc_inp->inp_socket->so_snd, 1026 BBR_LOG_PROGRESS, 0, 1027 0, &log, false); 1028 } 1029 } 1030 1031 static void 1032 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts) 1033 { 1034 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1035 union tcp_log_stackspecific log; 1036 1037 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1038 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1039 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1040 log.u_bbr.flex1 = slot; 1041 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); 1042 log.u_bbr.flex8 = rack->rc_in_persist; 1043 TCP_LOG_EVENT(rack->rc_tp, NULL, 1044 &rack->rc_inp->inp_socket->so_rcv, 1045 &rack->rc_inp->inp_socket->so_snd, 1046 BBR_LOG_BBRSND, 0, 1047 0, &log, false); 1048 } 1049 } 1050 1051 static void 1052 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) 1053 { 1054 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1055 union tcp_log_stackspecific log; 1056 1057 memset(&log, 0, sizeof(log)); 1058 log.u_bbr.flex1 = did_out; 1059 log.u_bbr.flex2 = nxt_pkt; 1060 log.u_bbr.flex3 = way_out; 1061 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1062 log.u_bbr.flex7 = rack->r_wanted_output; 1063 log.u_bbr.flex8 = rack->rc_in_persist; 1064 TCP_LOG_EVENT(rack->rc_tp, NULL, 1065 &rack->rc_inp->inp_socket->so_rcv, 1066 &rack->rc_inp->inp_socket->so_snd, 1067 BBR_LOG_DOSEG_DONE, 0, 1068 0, &log, false); 1069 } 1070 } 1071 1072 1073 static void 1074 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling) 1075 { 1076 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1077 union tcp_log_stackspecific log; 1078 1079 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1080 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1081 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1082 log.u_bbr.flex1 = slot; 1083 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; 1084 log.u_bbr.flex7 = hpts_calling; 1085 log.u_bbr.flex8 = rack->rc_in_persist; 1086 TCP_LOG_EVENT(rack->rc_tp, NULL, 1087 &rack->rc_inp->inp_socket->so_rcv, 1088 &rack->rc_inp->inp_socket->so_snd, 1089 BBR_LOG_JUSTRET, 0, 1090 tlen, &log, false); 1091 } 1092 } 1093 1094 static void 1095 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line) 1096 { 1097 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1098 union tcp_log_stackspecific log; 1099 1100 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1101 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 1102 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 1103 log.u_bbr.flex1 = line; 1104 log.u_bbr.flex2 = 0; 1105 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; 1106 log.u_bbr.flex4 = 0; 1107 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; 1108 log.u_bbr.flex8 = hpts_removed; 1109 TCP_LOG_EVENT(rack->rc_tp, NULL, 1110 &rack->rc_inp->inp_socket->so_rcv, 1111 &rack->rc_inp->inp_socket->so_snd, 1112 BBR_LOG_TIMERCANC, 0, 1113 0, &log, false); 1114 } 1115 } 1116 1117 static void 1118 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) 1119 { 1120 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1121 union tcp_log_stackspecific log; 1122 1123 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 1124 log.u_bbr.flex1 = timers; 1125 log.u_bbr.flex2 = ret; 1126 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; 1127 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; 1128 log.u_bbr.flex5 = cts; 1129 TCP_LOG_EVENT(rack->rc_tp, NULL, 1130 &rack->rc_inp->inp_socket->so_rcv, 1131 &rack->rc_inp->inp_socket->so_snd, 1132 BBR_LOG_TO_PROCESS, 0, 1133 0, &log, false); 1134 } 1135 } 1136 1137 static void 1138 rack_counter_destroy() 1139 { 1140 counter_u64_free(rack_badfr); 1141 counter_u64_free(rack_badfr_bytes); 1142 counter_u64_free(rack_rtm_prr_retran); 1143 counter_u64_free(rack_rtm_prr_newdata); 1144 counter_u64_free(rack_timestamp_mismatch); 1145 counter_u64_free(rack_reorder_seen); 1146 counter_u64_free(rack_tlp_tot); 1147 counter_u64_free(rack_tlp_newdata); 1148 counter_u64_free(rack_tlp_retran); 1149 counter_u64_free(rack_tlp_retran_bytes); 1150 counter_u64_free(rack_tlp_retran_fail); 1151 counter_u64_free(rack_to_tot); 1152 counter_u64_free(rack_to_arm_rack); 1153 counter_u64_free(rack_to_arm_tlp); 1154 counter_u64_free(rack_paced_segments); 1155 counter_u64_free(rack_unpaced_segments); 1156 counter_u64_free(rack_saw_enobuf); 1157 counter_u64_free(rack_saw_enetunreach); 1158 counter_u64_free(rack_to_alloc_hard); 1159 counter_u64_free(rack_to_alloc_emerg); 1160 counter_u64_free(rack_sack_proc_all); 1161 counter_u64_free(rack_sack_proc_short); 1162 counter_u64_free(rack_sack_proc_restart); 1163 counter_u64_free(rack_to_alloc); 1164 counter_u64_free(rack_to_alloc_limited); 1165 counter_u64_free(rack_split_limited); 1166 counter_u64_free(rack_find_high); 1167 counter_u64_free(rack_runt_sacks); 1168 counter_u64_free(rack_enter_tlp_calc); 1169 counter_u64_free(rack_used_tlpmethod); 1170 counter_u64_free(rack_used_tlpmethod2); 1171 counter_u64_free(rack_progress_drops); 1172 counter_u64_free(rack_input_idle_reduces); 1173 counter_u64_free(rack_tlp_does_nada); 1174 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); 1175 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); 1176 } 1177 1178 static struct rack_sendmap * 1179 rack_alloc(struct tcp_rack *rack) 1180 { 1181 struct rack_sendmap *rsm; 1182 1183 rsm = uma_zalloc(rack_zone, M_NOWAIT); 1184 if (rsm) { 1185 rack->r_ctl.rc_num_maps_alloced++; 1186 counter_u64_add(rack_to_alloc, 1); 1187 return (rsm); 1188 } 1189 if (rack->rc_free_cnt) { 1190 counter_u64_add(rack_to_alloc_emerg, 1); 1191 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 1192 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 1193 rack->rc_free_cnt--; 1194 return (rsm); 1195 } 1196 return (NULL); 1197 } 1198 1199 static struct rack_sendmap * 1200 rack_alloc_full_limit(struct tcp_rack *rack) 1201 { 1202 if ((rack_map_entries_limit > 0) && 1203 (rack->r_ctl.rc_num_maps_alloced >= rack_map_entries_limit)) { 1204 counter_u64_add(rack_to_alloc_limited, 1); 1205 if (!rack->alloc_limit_reported) { 1206 rack->alloc_limit_reported = 1; 1207 counter_u64_add(rack_alloc_limited_conns, 1); 1208 } 1209 return (NULL); 1210 } 1211 return (rack_alloc(rack)); 1212 } 1213 1214 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 1215 static struct rack_sendmap * 1216 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) 1217 { 1218 struct rack_sendmap *rsm; 1219 1220 if (limit_type) { 1221 /* currently there is only one limit type */ 1222 if (rack_map_split_limit > 0 && 1223 rack->r_ctl.rc_num_split_allocs >= rack_map_split_limit) { 1224 counter_u64_add(rack_split_limited, 1); 1225 if (!rack->alloc_limit_reported) { 1226 rack->alloc_limit_reported = 1; 1227 counter_u64_add(rack_alloc_limited_conns, 1); 1228 } 1229 return (NULL); 1230 } 1231 } 1232 1233 /* allocate and mark in the limit type, if set */ 1234 rsm = rack_alloc(rack); 1235 if (rsm != NULL && limit_type) { 1236 rsm->r_limit_type = limit_type; 1237 rack->r_ctl.rc_num_split_allocs++; 1238 } 1239 return (rsm); 1240 } 1241 1242 static void 1243 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) 1244 { 1245 if (rsm->r_limit_type) { 1246 /* currently there is only one limit type */ 1247 rack->r_ctl.rc_num_split_allocs--; 1248 } 1249 if (rack->r_ctl.rc_tlpsend == rsm) 1250 rack->r_ctl.rc_tlpsend = NULL; 1251 if (rack->r_ctl.rc_next == rsm) 1252 rack->r_ctl.rc_next = NULL; 1253 if (rack->r_ctl.rc_sacklast == rsm) 1254 rack->r_ctl.rc_sacklast = NULL; 1255 if (rack->rc_free_cnt < rack_free_cache) { 1256 memset(rsm, 0, sizeof(struct rack_sendmap)); 1257 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 1258 rsm->r_limit_type = 0; 1259 rack->rc_free_cnt++; 1260 return; 1261 } 1262 rack->r_ctl.rc_num_maps_alloced--; 1263 uma_zfree(rack_zone, rsm); 1264 } 1265 1266 /* 1267 * CC wrapper hook functions 1268 */ 1269 static void 1270 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, 1271 uint16_t type, int32_t recovery) 1272 { 1273 #ifdef NETFLIX_STATS 1274 int32_t gput; 1275 #endif 1276 1277 INP_WLOCK_ASSERT(tp->t_inpcb); 1278 1279 tp->ccv->nsegs = nsegs; 1280 tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); 1281 if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { 1282 uint32_t max; 1283 1284 max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg; 1285 if (tp->ccv->bytes_this_ack > max) { 1286 tp->ccv->bytes_this_ack = max; 1287 } 1288 } 1289 if (tp->snd_cwnd <= tp->snd_wnd) 1290 tp->ccv->flags |= CCF_CWND_LIMITED; 1291 else 1292 tp->ccv->flags &= ~CCF_CWND_LIMITED; 1293 1294 if (type == CC_ACK) { 1295 #ifdef NETFLIX_STATS 1296 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, 1297 ((int32_t) tp->snd_cwnd) - tp->snd_wnd); 1298 if ((tp->t_flags & TF_GPUTINPROG) && 1299 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 1300 gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / 1301 max(1, tcp_ts_getticks() - tp->gput_ts); 1302 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 1303 gput); 1304 /* 1305 * XXXLAS: This is a temporary hack, and should be 1306 * chained off VOI_TCP_GPUT when stats(9) grows an 1307 * API to deal with chained VOIs. 1308 */ 1309 if (tp->t_stats_gput_prev > 0) 1310 stats_voi_update_abs_s32(tp->t_stats, 1311 VOI_TCP_GPUT_ND, 1312 ((gput - tp->t_stats_gput_prev) * 100) / 1313 tp->t_stats_gput_prev); 1314 tp->t_flags &= ~TF_GPUTINPROG; 1315 tp->t_stats_gput_prev = gput; 1316 if (tp->t_maxpeakrate) { 1317 /* 1318 * We update t_peakrate_thr. This gives us roughly 1319 * one update per round trip time. 1320 */ 1321 tcp_update_peakrate_thr(tp); 1322 } 1323 } 1324 #endif 1325 if (tp->snd_cwnd > tp->snd_ssthresh) { 1326 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, 1327 nsegs * V_tcp_abc_l_var * tp->t_maxseg); 1328 if (tp->t_bytes_acked >= tp->snd_cwnd) { 1329 tp->t_bytes_acked -= tp->snd_cwnd; 1330 tp->ccv->flags |= CCF_ABC_SENTAWND; 1331 } 1332 } else { 1333 tp->ccv->flags &= ~CCF_ABC_SENTAWND; 1334 tp->t_bytes_acked = 0; 1335 } 1336 } 1337 if (CC_ALGO(tp)->ack_received != NULL) { 1338 /* XXXLAS: Find a way to live without this */ 1339 tp->ccv->curack = th->th_ack; 1340 CC_ALGO(tp)->ack_received(tp->ccv, type); 1341 } 1342 #ifdef NETFLIX_STATS 1343 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); 1344 #endif 1345 if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { 1346 rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; 1347 } 1348 /* we enforce max peak rate if it is set. */ 1349 if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { 1350 tp->snd_cwnd = tp->t_peakrate_thr; 1351 } 1352 } 1353 1354 static void 1355 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) 1356 { 1357 struct tcp_rack *rack; 1358 1359 rack = (struct tcp_rack *)tp->t_fb_ptr; 1360 INP_WLOCK_ASSERT(tp->t_inpcb); 1361 if (rack->r_ctl.rc_prr_sndcnt > 0) 1362 rack->r_wanted_output++; 1363 } 1364 1365 static void 1366 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) 1367 { 1368 struct tcp_rack *rack; 1369 1370 INP_WLOCK_ASSERT(tp->t_inpcb); 1371 rack = (struct tcp_rack *)tp->t_fb_ptr; 1372 if (CC_ALGO(tp)->post_recovery != NULL) { 1373 tp->ccv->curack = th->th_ack; 1374 CC_ALGO(tp)->post_recovery(tp->ccv); 1375 } 1376 /* 1377 * Here we can in theory adjust cwnd to be based on the number of 1378 * losses in the window (rack->r_ctl.rc_loss_count). This is done 1379 * based on the rack_use_proportional flag. 1380 */ 1381 if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { 1382 int32_t reduce; 1383 1384 reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); 1385 if (reduce > 50) { 1386 reduce = 50; 1387 } 1388 tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); 1389 } else { 1390 if (tp->snd_cwnd > tp->snd_ssthresh) { 1391 /* Drop us down to the ssthresh (1/2 cwnd at loss) */ 1392 tp->snd_cwnd = tp->snd_ssthresh; 1393 } 1394 } 1395 if (rack->r_ctl.rc_prr_sndcnt > 0) { 1396 /* Suck the next prr cnt back into cwnd */ 1397 tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; 1398 rack->r_ctl.rc_prr_sndcnt = 0; 1399 } 1400 tp->snd_recover = tp->snd_una; 1401 EXIT_RECOVERY(tp->t_flags); 1402 } 1403 1404 static void 1405 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) 1406 { 1407 struct tcp_rack *rack; 1408 1409 INP_WLOCK_ASSERT(tp->t_inpcb); 1410 1411 rack = (struct tcp_rack *)tp->t_fb_ptr; 1412 switch (type) { 1413 case CC_NDUPACK: 1414 /* rack->r_ctl.rc_ssthresh_set = 1;*/ 1415 if (!IN_FASTRECOVERY(tp->t_flags)) { 1416 rack->r_ctl.rc_tlp_rtx_out = 0; 1417 rack->r_ctl.rc_prr_delivered = 0; 1418 rack->r_ctl.rc_prr_out = 0; 1419 rack->r_ctl.rc_loss_count = 0; 1420 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 1421 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; 1422 tp->snd_recover = tp->snd_max; 1423 if (tp->t_flags & TF_ECN_PERMIT) 1424 tp->t_flags |= TF_ECN_SND_CWR; 1425 } 1426 break; 1427 case CC_ECN: 1428 if (!IN_CONGRECOVERY(tp->t_flags)) { 1429 TCPSTAT_INC(tcps_ecn_rcwnd); 1430 tp->snd_recover = tp->snd_max; 1431 if (tp->t_flags & TF_ECN_PERMIT) 1432 tp->t_flags |= TF_ECN_SND_CWR; 1433 } 1434 break; 1435 case CC_RTO: 1436 tp->t_dupacks = 0; 1437 tp->t_bytes_acked = 0; 1438 EXIT_RECOVERY(tp->t_flags); 1439 tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / 1440 tp->t_maxseg) * tp->t_maxseg; 1441 tp->snd_cwnd = tp->t_maxseg; 1442 break; 1443 case CC_RTO_ERR: 1444 TCPSTAT_INC(tcps_sndrexmitbad); 1445 /* RTO was unnecessary, so reset everything. */ 1446 tp->snd_cwnd = tp->snd_cwnd_prev; 1447 tp->snd_ssthresh = tp->snd_ssthresh_prev; 1448 tp->snd_recover = tp->snd_recover_prev; 1449 if (tp->t_flags & TF_WASFRECOVERY) 1450 ENTER_FASTRECOVERY(tp->t_flags); 1451 if (tp->t_flags & TF_WASCRECOVERY) 1452 ENTER_CONGRECOVERY(tp->t_flags); 1453 tp->snd_nxt = tp->snd_max; 1454 tp->t_badrxtwin = 0; 1455 break; 1456 } 1457 1458 if (CC_ALGO(tp)->cong_signal != NULL) { 1459 if (th != NULL) 1460 tp->ccv->curack = th->th_ack; 1461 CC_ALGO(tp)->cong_signal(tp->ccv, type); 1462 } 1463 } 1464 1465 1466 1467 static inline void 1468 rack_cc_after_idle(struct tcpcb *tp, int reduce_largest) 1469 { 1470 uint32_t i_cwnd; 1471 1472 INP_WLOCK_ASSERT(tp->t_inpcb); 1473 1474 #ifdef NETFLIX_STATS 1475 TCPSTAT_INC(tcps_idle_restarts); 1476 if (tp->t_state == TCPS_ESTABLISHED) 1477 TCPSTAT_INC(tcps_idle_estrestarts); 1478 #endif 1479 if (CC_ALGO(tp)->after_idle != NULL) 1480 CC_ALGO(tp)->after_idle(tp->ccv); 1481 1482 if (V_tcp_initcwnd_segments) 1483 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 1484 max(2 * tp->t_maxseg, 14600)); 1485 else if (V_tcp_do_rfc3390) 1486 i_cwnd = min(4 * tp->t_maxseg, 1487 max(2 * tp->t_maxseg, 4380)); 1488 else { 1489 /* Per RFC5681 Section 3.1 */ 1490 if (tp->t_maxseg > 2190) 1491 i_cwnd = 2 * tp->t_maxseg; 1492 else if (tp->t_maxseg > 1095) 1493 i_cwnd = 3 * tp->t_maxseg; 1494 else 1495 i_cwnd = 4 * tp->t_maxseg; 1496 } 1497 if (reduce_largest) { 1498 /* 1499 * Do we reduce the largest cwnd to make 1500 * rack play nice on restart hptsi wise? 1501 */ 1502 if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) 1503 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; 1504 } 1505 /* 1506 * Being idle is no differnt than the initial window. If the cc 1507 * clamps it down below the initial window raise it to the initial 1508 * window. 1509 */ 1510 if (tp->snd_cwnd < i_cwnd) { 1511 tp->snd_cwnd = i_cwnd; 1512 } 1513 } 1514 1515 1516 /* 1517 * Indicate whether this ack should be delayed. We can delay the ack if 1518 * following conditions are met: 1519 * - There is no delayed ack timer in progress. 1520 * - Our last ack wasn't a 0-sized window. We never want to delay 1521 * the ack that opens up a 0-sized window. 1522 * - LRO wasn't used for this segment. We make sure by checking that the 1523 * segment size is not larger than the MSS. 1524 * - Delayed acks are enabled or this is a half-synchronized T/TCP 1525 * connection. 1526 */ 1527 #define DELAY_ACK(tp, tlen) \ 1528 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 1529 ((tp->t_flags & TF_DELACK) == 0) && \ 1530 (tlen <= tp->t_maxseg) && \ 1531 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 1532 1533 static inline void 1534 rack_calc_rwin(struct socket *so, struct tcpcb *tp) 1535 { 1536 int32_t win; 1537 1538 /* 1539 * Calculate amount of space in receive window, and then do TCP 1540 * input processing. Receive window is amount of space in rcv queue, 1541 * but not less than advertised window. 1542 */ 1543 win = sbspace(&so->so_rcv); 1544 if (win < 0) 1545 win = 0; 1546 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 1547 } 1548 1549 static void 1550 rack_do_drop(struct mbuf *m, struct tcpcb *tp) 1551 { 1552 /* 1553 * Drop space held by incoming segment and return. 1554 */ 1555 if (tp != NULL) 1556 INP_WUNLOCK(tp->t_inpcb); 1557 if (m) 1558 m_freem(m); 1559 } 1560 1561 static void 1562 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t rstreason, int32_t tlen) 1563 { 1564 if (tp != NULL) { 1565 tcp_dropwithreset(m, th, tp, tlen, rstreason); 1566 INP_WUNLOCK(tp->t_inpcb); 1567 } else 1568 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 1569 } 1570 1571 /* 1572 * The value in ret_val informs the caller 1573 * if we dropped the tcb (and lock) or not. 1574 * 1 = we dropped it, 0 = the TCB is still locked 1575 * and valid. 1576 */ 1577 static void 1578 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val) 1579 { 1580 /* 1581 * Generate an ACK dropping incoming segment if it occupies sequence 1582 * space, where the ACK reflects our state. 1583 * 1584 * We can now skip the test for the RST flag since all paths to this 1585 * code happen after packets containing RST have been dropped. 1586 * 1587 * In the SYN-RECEIVED state, don't send an ACK unless the segment 1588 * we received passes the SYN-RECEIVED ACK test. If it fails send a 1589 * RST. This breaks the loop in the "LAND" DoS attack, and also 1590 * prevents an ACK storm between two listening ports that have been 1591 * sent forged SYN segments, each with the source address of the 1592 * other. 1593 */ 1594 struct tcp_rack *rack; 1595 1596 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 1597 (SEQ_GT(tp->snd_una, th->th_ack) || 1598 SEQ_GT(th->th_ack, tp->snd_max))) { 1599 *ret_val = 1; 1600 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 1601 return; 1602 } else 1603 *ret_val = 0; 1604 rack = (struct tcp_rack *)tp->t_fb_ptr; 1605 rack->r_wanted_output++; 1606 tp->t_flags |= TF_ACKNOW; 1607 if (m) 1608 m_freem(m); 1609 } 1610 1611 1612 static int 1613 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 1614 { 1615 /* 1616 * RFC5961 Section 3.2 1617 * 1618 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 1619 * window, we send challenge ACK. 1620 * 1621 * Note: to take into account delayed ACKs, we should test against 1622 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 1623 * of closed window, not covered by the RFC. 1624 */ 1625 int dropped = 0; 1626 1627 if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && 1628 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 1629 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 1630 1631 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1632 KASSERT(tp->t_state != TCPS_SYN_SENT, 1633 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 1634 __func__, th, tp)); 1635 1636 if (V_tcp_insecure_rst || 1637 (tp->last_ack_sent == th->th_seq) || 1638 (tp->rcv_nxt == th->th_seq) || 1639 ((tp->last_ack_sent - 1) == th->th_seq)) { 1640 TCPSTAT_INC(tcps_drops); 1641 /* Drop the connection. */ 1642 switch (tp->t_state) { 1643 case TCPS_SYN_RECEIVED: 1644 so->so_error = ECONNREFUSED; 1645 goto close; 1646 case TCPS_ESTABLISHED: 1647 case TCPS_FIN_WAIT_1: 1648 case TCPS_FIN_WAIT_2: 1649 case TCPS_CLOSE_WAIT: 1650 case TCPS_CLOSING: 1651 case TCPS_LAST_ACK: 1652 so->so_error = ECONNRESET; 1653 close: 1654 tcp_state_change(tp, TCPS_CLOSED); 1655 /* FALLTHROUGH */ 1656 default: 1657 tp = tcp_close(tp); 1658 } 1659 dropped = 1; 1660 rack_do_drop(m, tp); 1661 } else { 1662 TCPSTAT_INC(tcps_badrst); 1663 /* Send challenge ACK. */ 1664 tcp_respond(tp, mtod(m, void *), th, m, 1665 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 1666 tp->last_ack_sent = tp->rcv_nxt; 1667 } 1668 } else { 1669 m_freem(m); 1670 } 1671 return (dropped); 1672 } 1673 1674 /* 1675 * The value in ret_val informs the caller 1676 * if we dropped the tcb (and lock) or not. 1677 * 1 = we dropped it, 0 = the TCB is still locked 1678 * and valid. 1679 */ 1680 static void 1681 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 1682 { 1683 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 1684 1685 TCPSTAT_INC(tcps_badsyn); 1686 if (V_tcp_insecure_syn && 1687 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 1688 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 1689 tp = tcp_drop(tp, ECONNRESET); 1690 *ret_val = 1; 1691 rack_do_drop(m, tp); 1692 } else { 1693 /* Send challenge ACK. */ 1694 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 1695 tp->snd_nxt, TH_ACK); 1696 tp->last_ack_sent = tp->rcv_nxt; 1697 m = NULL; 1698 *ret_val = 0; 1699 rack_do_drop(m, NULL); 1700 } 1701 } 1702 1703 /* 1704 * rack_ts_check returns 1 for you should not proceed. It places 1705 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1706 * that the TCB is unlocked and probably dropped. The 0 indicates the 1707 * TCB is still valid and locked. 1708 */ 1709 static int 1710 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val) 1711 { 1712 1713 /* Check to see if ts_recent is over 24 days old. */ 1714 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 1715 /* 1716 * Invalidate ts_recent. If this segment updates ts_recent, 1717 * the age will be reset later and ts_recent will get a 1718 * valid value. If it does not, setting ts_recent to zero 1719 * will at least satisfy the requirement that zero be placed 1720 * in the timestamp echo reply when ts_recent isn't valid. 1721 * The age isn't reset until we get a valid ts_recent 1722 * because we don't want out-of-order segments to be dropped 1723 * when ts_recent is old. 1724 */ 1725 tp->ts_recent = 0; 1726 } else { 1727 TCPSTAT_INC(tcps_rcvduppack); 1728 TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 1729 TCPSTAT_INC(tcps_pawsdrop); 1730 *ret_val = 0; 1731 if (tlen) { 1732 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1733 } else { 1734 rack_do_drop(m, NULL); 1735 } 1736 return (1); 1737 } 1738 return (0); 1739 } 1740 1741 /* 1742 * rack_drop_checks returns 1 for you should not proceed. It places 1743 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 1744 * that the TCB is unlocked and probably dropped. The 0 indicates the 1745 * TCB is still valid and locked. 1746 */ 1747 static int 1748 rack_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val) 1749 { 1750 int32_t todrop; 1751 int32_t thflags; 1752 int32_t tlen; 1753 1754 thflags = *thf; 1755 tlen = *tlenp; 1756 todrop = tp->rcv_nxt - th->th_seq; 1757 if (todrop > 0) { 1758 if (thflags & TH_SYN) { 1759 thflags &= ~TH_SYN; 1760 th->th_seq++; 1761 if (th->th_urp > 1) 1762 th->th_urp--; 1763 else 1764 thflags &= ~TH_URG; 1765 todrop--; 1766 } 1767 /* 1768 * Following if statement from Stevens, vol. 2, p. 960. 1769 */ 1770 if (todrop > tlen 1771 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 1772 /* 1773 * Any valid FIN must be to the left of the window. 1774 * At this point the FIN must be a duplicate or out 1775 * of sequence; drop it. 1776 */ 1777 thflags &= ~TH_FIN; 1778 /* 1779 * Send an ACK to resynchronize and drop any data. 1780 * But keep on processing for RST or ACK. 1781 */ 1782 tp->t_flags |= TF_ACKNOW; 1783 todrop = tlen; 1784 TCPSTAT_INC(tcps_rcvduppack); 1785 TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 1786 } else { 1787 TCPSTAT_INC(tcps_rcvpartduppack); 1788 TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 1789 } 1790 *drop_hdrlen += todrop; /* drop from the top afterwards */ 1791 th->th_seq += todrop; 1792 tlen -= todrop; 1793 if (th->th_urp > todrop) 1794 th->th_urp -= todrop; 1795 else { 1796 thflags &= ~TH_URG; 1797 th->th_urp = 0; 1798 } 1799 } 1800 /* 1801 * If segment ends after window, drop trailing data (and PUSH and 1802 * FIN); if nothing left, just ACK. 1803 */ 1804 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 1805 if (todrop > 0) { 1806 TCPSTAT_INC(tcps_rcvpackafterwin); 1807 if (todrop >= tlen) { 1808 TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 1809 /* 1810 * If window is closed can only take segments at 1811 * window edge, and have to drop data and PUSH from 1812 * incoming segments. Continue processing, but 1813 * remember to ack. Otherwise, drop segment and 1814 * ack. 1815 */ 1816 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 1817 tp->t_flags |= TF_ACKNOW; 1818 TCPSTAT_INC(tcps_rcvwinprobe); 1819 } else { 1820 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 1821 return (1); 1822 } 1823 } else 1824 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 1825 m_adj(m, -todrop); 1826 tlen -= todrop; 1827 thflags &= ~(TH_PUSH | TH_FIN); 1828 } 1829 *thf = thflags; 1830 *tlenp = tlen; 1831 return (0); 1832 } 1833 1834 static struct rack_sendmap * 1835 rack_find_lowest_rsm(struct tcp_rack *rack) 1836 { 1837 struct rack_sendmap *rsm; 1838 1839 /* 1840 * Walk the time-order transmitted list looking for an rsm that is 1841 * not acked. This will be the one that was sent the longest time 1842 * ago that is still outstanding. 1843 */ 1844 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { 1845 if (rsm->r_flags & RACK_ACKED) { 1846 continue; 1847 } 1848 goto finish; 1849 } 1850 finish: 1851 return (rsm); 1852 } 1853 1854 static struct rack_sendmap * 1855 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) 1856 { 1857 struct rack_sendmap *prsm; 1858 1859 /* 1860 * Walk the sequence order list backward until we hit and arrive at 1861 * the highest seq not acked. In theory when this is called it 1862 * should be the last segment (which it was not). 1863 */ 1864 counter_u64_add(rack_find_high, 1); 1865 prsm = rsm; 1866 TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { 1867 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { 1868 continue; 1869 } 1870 return (prsm); 1871 } 1872 return (NULL); 1873 } 1874 1875 1876 static uint32_t 1877 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) 1878 { 1879 int32_t lro; 1880 uint32_t thresh; 1881 1882 /* 1883 * lro is the flag we use to determine if we have seen reordering. 1884 * If it gets set we have seen reordering. The reorder logic either 1885 * works in one of two ways: 1886 * 1887 * If reorder-fade is configured, then we track the last time we saw 1888 * re-ordering occur. If we reach the point where enough time as 1889 * passed we no longer consider reordering has occuring. 1890 * 1891 * Or if reorder-face is 0, then once we see reordering we consider 1892 * the connection to alway be subject to reordering and just set lro 1893 * to 1. 1894 * 1895 * In the end if lro is non-zero we add the extra time for 1896 * reordering in. 1897 */ 1898 if (srtt == 0) 1899 srtt = 1; 1900 if (rack->r_ctl.rc_reorder_ts) { 1901 if (rack->r_ctl.rc_reorder_fade) { 1902 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { 1903 lro = cts - rack->r_ctl.rc_reorder_ts; 1904 if (lro == 0) { 1905 /* 1906 * No time as passed since the last 1907 * reorder, mark it as reordering. 1908 */ 1909 lro = 1; 1910 } 1911 } else { 1912 /* Negative time? */ 1913 lro = 0; 1914 } 1915 if (lro > rack->r_ctl.rc_reorder_fade) { 1916 /* Turn off reordering seen too */ 1917 rack->r_ctl.rc_reorder_ts = 0; 1918 lro = 0; 1919 } 1920 } else { 1921 /* Reodering does not fade */ 1922 lro = 1; 1923 } 1924 } else { 1925 lro = 0; 1926 } 1927 thresh = srtt + rack->r_ctl.rc_pkt_delay; 1928 if (lro) { 1929 /* It must be set, if not you get 1/4 rtt */ 1930 if (rack->r_ctl.rc_reorder_shift) 1931 thresh += (srtt >> rack->r_ctl.rc_reorder_shift); 1932 else 1933 thresh += (srtt >> 2); 1934 } else { 1935 thresh += 1; 1936 } 1937 /* We don't let the rack timeout be above a RTO */ 1938 1939 if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { 1940 thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); 1941 } 1942 /* And we don't want it above the RTO max either */ 1943 if (thresh > rack_rto_max) { 1944 thresh = rack_rto_max; 1945 } 1946 return (thresh); 1947 } 1948 1949 static uint32_t 1950 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, 1951 struct rack_sendmap *rsm, uint32_t srtt) 1952 { 1953 struct rack_sendmap *prsm; 1954 uint32_t thresh, len; 1955 int maxseg; 1956 1957 if (srtt == 0) 1958 srtt = 1; 1959 if (rack->r_ctl.rc_tlp_threshold) 1960 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); 1961 else 1962 thresh = (srtt * 2); 1963 1964 /* Get the previous sent packet, if any */ 1965 maxseg = tcp_maxseg(tp); 1966 counter_u64_add(rack_enter_tlp_calc, 1); 1967 len = rsm->r_end - rsm->r_start; 1968 if (rack->rack_tlp_threshold_use == TLP_USE_ID) { 1969 /* Exactly like the ID */ 1970 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { 1971 uint32_t alt_thresh; 1972 /* 1973 * Compensate for delayed-ack with the d-ack time. 1974 */ 1975 counter_u64_add(rack_used_tlpmethod, 1); 1976 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 1977 if (alt_thresh > thresh) 1978 thresh = alt_thresh; 1979 } 1980 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { 1981 /* 2.1 behavior */ 1982 prsm = TAILQ_PREV(rsm, rack_head, r_tnext); 1983 if (prsm && (len <= maxseg)) { 1984 /* 1985 * Two packets outstanding, thresh should be (2*srtt) + 1986 * possible inter-packet delay (if any). 1987 */ 1988 uint32_t inter_gap = 0; 1989 int idx, nidx; 1990 1991 counter_u64_add(rack_used_tlpmethod, 1); 1992 idx = rsm->r_rtr_cnt - 1; 1993 nidx = prsm->r_rtr_cnt - 1; 1994 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 1995 /* Yes it was sent later (or at the same time) */ 1996 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 1997 } 1998 thresh += inter_gap; 1999 } else if (len <= maxseg) { 2000 /* 2001 * Possibly compensate for delayed-ack. 2002 */ 2003 uint32_t alt_thresh; 2004 2005 counter_u64_add(rack_used_tlpmethod2, 1); 2006 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2007 if (alt_thresh > thresh) 2008 thresh = alt_thresh; 2009 } 2010 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { 2011 /* 2.2 behavior */ 2012 if (len <= maxseg) { 2013 uint32_t alt_thresh; 2014 /* 2015 * Compensate for delayed-ack with the d-ack time. 2016 */ 2017 counter_u64_add(rack_used_tlpmethod, 1); 2018 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; 2019 if (alt_thresh > thresh) 2020 thresh = alt_thresh; 2021 } 2022 } 2023 /* Not above an RTO */ 2024 if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { 2025 thresh = TICKS_2_MSEC(tp->t_rxtcur); 2026 } 2027 /* Not above a RTO max */ 2028 if (thresh > rack_rto_max) { 2029 thresh = rack_rto_max; 2030 } 2031 /* Apply user supplied min TLP */ 2032 if (thresh < rack_tlp_min) { 2033 thresh = rack_tlp_min; 2034 } 2035 return (thresh); 2036 } 2037 2038 static struct rack_sendmap * 2039 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) 2040 { 2041 /* 2042 * Check to see that we don't need to fall into recovery. We will 2043 * need to do so if our oldest transmit is past the time we should 2044 * have had an ack. 2045 */ 2046 struct tcp_rack *rack; 2047 struct rack_sendmap *rsm; 2048 int32_t idx; 2049 uint32_t srtt_cur, srtt, thresh; 2050 2051 rack = (struct tcp_rack *)tp->t_fb_ptr; 2052 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 2053 return (NULL); 2054 } 2055 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 2056 srtt = TICKS_2_MSEC(srtt_cur); 2057 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 2058 srtt = rack->rc_rack_rtt; 2059 2060 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2061 if (rsm == NULL) 2062 return (NULL); 2063 2064 if (rsm->r_flags & RACK_ACKED) { 2065 rsm = rack_find_lowest_rsm(rack); 2066 if (rsm == NULL) 2067 return (NULL); 2068 } 2069 idx = rsm->r_rtr_cnt - 1; 2070 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 2071 if (tsused < rsm->r_tim_lastsent[idx]) { 2072 return (NULL); 2073 } 2074 if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { 2075 return (NULL); 2076 } 2077 /* Ok if we reach here we are over-due */ 2078 rack->r_ctl.rc_rsm_start = rsm->r_start; 2079 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 2080 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 2081 rack_cong_signal(tp, NULL, CC_NDUPACK); 2082 return (rsm); 2083 } 2084 2085 static uint32_t 2086 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) 2087 { 2088 int32_t t; 2089 int32_t tt; 2090 uint32_t ret_val; 2091 2092 t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); 2093 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], 2094 tcp_persmin, tcp_persmax); 2095 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 2096 tp->t_rxtshift++; 2097 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 2098 ret_val = (uint32_t)tt; 2099 return (ret_val); 2100 } 2101 2102 static uint32_t 2103 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2104 { 2105 /* 2106 * Start the FR timer, we do this based on getting the first one in 2107 * the rc_tmap. Note that if its NULL we must stop the timer. in all 2108 * events we need to stop the running timer (if its running) before 2109 * starting the new one. 2110 */ 2111 uint32_t thresh, exp, to, srtt, time_since_sent; 2112 uint32_t srtt_cur; 2113 int32_t idx; 2114 int32_t is_tlp_timer = 0; 2115 struct rack_sendmap *rsm; 2116 2117 if (rack->t_timers_stopped) { 2118 /* All timers have been stopped none are to run */ 2119 return (0); 2120 } 2121 if (rack->rc_in_persist) { 2122 /* We can't start any timer in persists */ 2123 return (rack_get_persists_timer_val(tp, rack)); 2124 } 2125 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2126 if (rsm == NULL) { 2127 /* Nothing on the send map */ 2128 activate_rxt: 2129 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 2130 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 2131 to = TICKS_2_MSEC(tp->t_rxtcur); 2132 if (to == 0) 2133 to = 1; 2134 return (to); 2135 } 2136 return (0); 2137 } 2138 if (rsm->r_flags & RACK_ACKED) { 2139 rsm = rack_find_lowest_rsm(rack); 2140 if (rsm == NULL) { 2141 /* No lowest? */ 2142 goto activate_rxt; 2143 } 2144 } 2145 /* Convert from ms to usecs */ 2146 if (rsm->r_flags & RACK_SACK_PASSED) { 2147 if ((tp->t_flags & TF_SENTFIN) && 2148 ((tp->snd_max - tp->snd_una) == 1) && 2149 (rsm->r_flags & RACK_HAS_FIN)) { 2150 /* 2151 * We don't start a rack timer if all we have is a 2152 * FIN outstanding. 2153 */ 2154 goto activate_rxt; 2155 } 2156 if (tp->t_srtt) { 2157 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2158 srtt = TICKS_2_MSEC(srtt_cur); 2159 } else 2160 srtt = RACK_INITIAL_RTO; 2161 2162 thresh = rack_calc_thresh_rack(rack, srtt, cts); 2163 idx = rsm->r_rtr_cnt - 1; 2164 exp = rsm->r_tim_lastsent[idx] + thresh; 2165 if (SEQ_GEQ(exp, cts)) { 2166 to = exp - cts; 2167 if (to < rack->r_ctl.rc_min_to) { 2168 to = rack->r_ctl.rc_min_to; 2169 } 2170 } else { 2171 to = rack->r_ctl.rc_min_to; 2172 } 2173 } else { 2174 /* Ok we need to do a TLP not RACK */ 2175 if ((rack->rc_tlp_in_progress != 0) || 2176 (rack->r_ctl.rc_tlp_rtx_out != 0)) { 2177 /* 2178 * The previous send was a TLP or a tlp_rtx is in 2179 * process. 2180 */ 2181 goto activate_rxt; 2182 } 2183 if ((tp->snd_max - tp->snd_una) > tp->snd_wnd) { 2184 /* 2185 * Peer collapsed rwnd, don't do TLP. 2186 */ 2187 goto activate_rxt; 2188 } 2189 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 2190 if (rsm == NULL) { 2191 /* We found no rsm to TLP with. */ 2192 goto activate_rxt; 2193 } 2194 if (rsm->r_flags & RACK_HAS_FIN) { 2195 /* If its a FIN we dont do TLP */ 2196 rsm = NULL; 2197 goto activate_rxt; 2198 } 2199 idx = rsm->r_rtr_cnt - 1; 2200 if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) 2201 time_since_sent = cts - rsm->r_tim_lastsent[idx]; 2202 else 2203 time_since_sent = 0; 2204 is_tlp_timer = 1; 2205 if (tp->t_srtt) { 2206 srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); 2207 srtt = TICKS_2_MSEC(srtt_cur); 2208 } else 2209 srtt = RACK_INITIAL_RTO; 2210 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); 2211 if (thresh > time_since_sent) 2212 to = thresh - time_since_sent; 2213 else 2214 to = rack->r_ctl.rc_min_to; 2215 if (to > TCPTV_REXMTMAX) { 2216 /* 2217 * If the TLP time works out to larger than the max 2218 * RTO lets not do TLP.. just RTO. 2219 */ 2220 goto activate_rxt; 2221 } 2222 if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { 2223 /* 2224 * The tail is no longer the last one I did a probe 2225 * on 2226 */ 2227 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2228 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2229 } 2230 } 2231 if (is_tlp_timer == 0) { 2232 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 2233 } else { 2234 if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || 2235 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2236 /* 2237 * We have exceeded how many times we can retran the 2238 * current TLP timer, switch to the RTO timer. 2239 */ 2240 goto activate_rxt; 2241 } else { 2242 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 2243 } 2244 } 2245 if (to == 0) 2246 to = 1; 2247 return (to); 2248 } 2249 2250 static void 2251 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2252 { 2253 if (rack->rc_in_persist == 0) { 2254 if (((tp->t_flags & TF_SENTFIN) == 0) && 2255 (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) 2256 /* Must need to send more data to enter persist */ 2257 return; 2258 rack->r_ctl.rc_went_idle_time = cts; 2259 rack_timer_cancel(tp, rack, cts, __LINE__); 2260 tp->t_rxtshift = 0; 2261 rack->rc_in_persist = 1; 2262 } 2263 } 2264 2265 static void 2266 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) 2267 { 2268 if (rack->rc_inp->inp_in_hpts) { 2269 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 2270 rack->r_ctl.rc_hpts_flags = 0; 2271 } 2272 rack->rc_in_persist = 0; 2273 rack->r_ctl.rc_went_idle_time = 0; 2274 tp->t_flags &= ~TF_FORCEDATA; 2275 tp->t_rxtshift = 0; 2276 } 2277 2278 static void 2279 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, 2280 int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) 2281 { 2282 struct inpcb *inp; 2283 uint32_t delayed_ack = 0; 2284 uint32_t hpts_timeout; 2285 uint8_t stopped; 2286 uint32_t left = 0; 2287 2288 inp = tp->t_inpcb; 2289 if (inp->inp_in_hpts) { 2290 /* A previous call is already set up */ 2291 return; 2292 } 2293 2294 if ((tp->t_state == TCPS_CLOSED) || 2295 (tp->t_state == TCPS_LISTEN)) { 2296 return; 2297 } 2298 stopped = rack->rc_tmr_stopped; 2299 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { 2300 left = rack->r_ctl.rc_timer_exp - cts; 2301 } 2302 rack->r_ctl.rc_timer_exp = 0; 2303 if (rack->rc_inp->inp_in_hpts == 0) { 2304 rack->r_ctl.rc_hpts_flags = 0; 2305 } 2306 if (slot) { 2307 /* We are hptsi too */ 2308 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 2309 } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 2310 /* 2311 * We are still left on the hpts when the to goes 2312 * it will be for output. 2313 */ 2314 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) 2315 slot = rack->r_ctl.rc_last_output_to - cts; 2316 else 2317 slot = 1; 2318 } 2319 if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { 2320 /* No send window.. we must enter persist */ 2321 rack_enter_persist(tp, rack, cts); 2322 } else if ((frm_out_sbavail && 2323 (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && 2324 (tp->snd_wnd < tp->t_maxseg)) && 2325 TCPS_HAVEESTABLISHED(tp->t_state)) { 2326 /* 2327 * If we have no window or we can't send a segment (and have 2328 * data to send.. we cheat here and frm_out_sbavail is 2329 * passed in with the sbavail(sb) only from bbr_output) and 2330 * we are established, then we must enter persits (if not 2331 * already in persits). 2332 */ 2333 rack_enter_persist(tp, rack, cts); 2334 } 2335 hpts_timeout = rack_timer_start(tp, rack, cts); 2336 if (tp->t_flags & TF_DELACK) { 2337 delayed_ack = tcp_delacktime; 2338 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; 2339 } 2340 if (delayed_ack && ((hpts_timeout == 0) || 2341 (delayed_ack < hpts_timeout))) 2342 hpts_timeout = delayed_ack; 2343 else 2344 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2345 /* 2346 * If no timers are going to run and we will fall off the hptsi 2347 * wheel, we resort to a keep-alive timer if its configured. 2348 */ 2349 if ((hpts_timeout == 0) && 2350 (slot == 0)) { 2351 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2352 (tp->t_state <= TCPS_CLOSING)) { 2353 /* 2354 * Ok we have no timer (persists, rack, tlp, rxt or 2355 * del-ack), we don't have segments being paced. So 2356 * all that is left is the keepalive timer. 2357 */ 2358 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 2359 /* Get the established keep-alive time */ 2360 hpts_timeout = TP_KEEPIDLE(tp); 2361 } else { 2362 /* Get the initial setup keep-alive time */ 2363 hpts_timeout = TP_KEEPINIT(tp); 2364 } 2365 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 2366 } 2367 } 2368 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 2369 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 2370 /* 2371 * RACK, TLP, persists and RXT timers all are restartable 2372 * based on actions input .. i.e we received a packet (ack 2373 * or sack) and that changes things (rw, or snd_una etc). 2374 * Thus we can restart them with a new value. For 2375 * keep-alive, delayed_ack we keep track of what was left 2376 * and restart the timer with a smaller value. 2377 */ 2378 if (left < hpts_timeout) 2379 hpts_timeout = left; 2380 } 2381 if (hpts_timeout) { 2382 /* 2383 * Hack alert for now we can't time-out over 2,147,483 2384 * seconds (a bit more than 596 hours), which is probably ok 2385 * :). 2386 */ 2387 if (hpts_timeout > 0x7ffffffe) 2388 hpts_timeout = 0x7ffffffe; 2389 rack->r_ctl.rc_timer_exp = cts + hpts_timeout; 2390 } 2391 if (slot) { 2392 rack->r_ctl.rc_last_output_to = cts + slot; 2393 if ((hpts_timeout == 0) || (hpts_timeout > slot)) { 2394 if (rack->rc_inp->inp_in_hpts == 0) 2395 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); 2396 rack_log_to_start(rack, cts, hpts_timeout, slot, 1); 2397 } else { 2398 /* 2399 * Arrange for the hpts to kick back in after the 2400 * t-o if the t-o does not cause a send. 2401 */ 2402 if (rack->rc_inp->inp_in_hpts == 0) 2403 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2404 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2405 } 2406 } else if (hpts_timeout) { 2407 if (rack->rc_inp->inp_in_hpts == 0) 2408 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 2409 rack_log_to_start(rack, cts, hpts_timeout, slot, 0); 2410 } else { 2411 /* No timer starting */ 2412 #ifdef INVARIANTS 2413 if (SEQ_GT(tp->snd_max, tp->snd_una)) { 2414 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", 2415 tp, rack, tot_len_this_send, cts, slot, hpts_timeout); 2416 } 2417 #endif 2418 } 2419 rack->rc_tmr_stopped = 0; 2420 if (slot) 2421 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); 2422 } 2423 2424 /* 2425 * RACK Timer, here we simply do logging and house keeping. 2426 * the normal rack_output() function will call the 2427 * appropriate thing to check if we need to do a RACK retransmit. 2428 * We return 1, saying don't proceed with rack_output only 2429 * when all timers have been stopped (destroyed PCB?). 2430 */ 2431 static int 2432 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2433 { 2434 /* 2435 * This timer simply provides an internal trigger to send out data. 2436 * The check_recovery_mode call will see if there are needed 2437 * retransmissions, if so we will enter fast-recovery. The output 2438 * call may or may not do the same thing depending on sysctl 2439 * settings. 2440 */ 2441 struct rack_sendmap *rsm; 2442 int32_t recovery; 2443 2444 if (tp->t_timers->tt_flags & TT_STOPPED) { 2445 return (1); 2446 } 2447 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2448 /* Its not time yet */ 2449 return (0); 2450 } 2451 rack_log_to_event(rack, RACK_TO_FRM_RACK); 2452 recovery = IN_RECOVERY(tp->t_flags); 2453 counter_u64_add(rack_to_tot, 1); 2454 if (rack->r_state && (rack->r_state != tp->t_state)) 2455 rack_set_state(tp, rack); 2456 rsm = rack_check_recovery_mode(tp, cts); 2457 if (rsm) { 2458 uint32_t rtt; 2459 2460 rtt = rack->rc_rack_rtt; 2461 if (rtt == 0) 2462 rtt = 1; 2463 if ((recovery == 0) && 2464 (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { 2465 /* 2466 * The rack-timeout that enter's us into recovery 2467 * will force out one MSS and set us up so that we 2468 * can do one more send in 2*rtt (transitioning the 2469 * rack timeout into a rack-tlp). 2470 */ 2471 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2472 } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && 2473 ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { 2474 /* 2475 * When a rack timer goes, we have to send at 2476 * least one segment. They will be paced a min of 1ms 2477 * apart via the next rack timer (or further 2478 * if the rack timer dictates it). 2479 */ 2480 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2481 } 2482 } else { 2483 /* This is a case that should happen rarely if ever */ 2484 counter_u64_add(rack_tlp_does_nada, 1); 2485 #ifdef TCP_BLACKBOX 2486 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2487 #endif 2488 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2489 } 2490 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 2491 return (0); 2492 } 2493 2494 static struct rack_sendmap * 2495 rack_merge_rsm(struct tcp_rack *rack, 2496 struct rack_sendmap *l_rsm, 2497 struct rack_sendmap *r_rsm) 2498 { 2499 /* 2500 * We are merging two ack'd RSM's, 2501 * the l_rsm is on the left (lower seq 2502 * values) and the r_rsm is on the right 2503 * (higher seq value). The simplest way 2504 * to merge these is to move the right 2505 * one into the left. I don't think there 2506 * is any reason we need to try to find 2507 * the oldest (or last oldest retransmitted). 2508 */ 2509 l_rsm->r_end = r_rsm->r_end; 2510 if (r_rsm->r_rtr_bytes) 2511 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 2512 if (r_rsm->r_in_tmap) { 2513 /* This really should not happen */ 2514 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext); 2515 } 2516 /* Now the flags */ 2517 if (r_rsm->r_flags & RACK_HAS_FIN) 2518 l_rsm->r_flags |= RACK_HAS_FIN; 2519 if (r_rsm->r_flags & RACK_TLP) 2520 l_rsm->r_flags |= RACK_TLP; 2521 TAILQ_REMOVE(&rack->r_ctl.rc_map, r_rsm, r_next); 2522 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 2523 /* Transfer the split limit to the map we free */ 2524 r_rsm->r_limit_type = l_rsm->r_limit_type; 2525 l_rsm->r_limit_type = 0; 2526 } 2527 rack_free(rack, r_rsm); 2528 return(l_rsm); 2529 } 2530 2531 /* 2532 * TLP Timer, here we simply setup what segment we want to 2533 * have the TLP expire on, the normal rack_output() will then 2534 * send it out. 2535 * 2536 * We return 1, saying don't proceed with rack_output only 2537 * when all timers have been stopped (destroyed PCB?). 2538 */ 2539 static int 2540 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2541 { 2542 /* 2543 * Tail Loss Probe. 2544 */ 2545 struct rack_sendmap *rsm = NULL; 2546 struct socket *so; 2547 uint32_t amm, old_prr_snd = 0; 2548 uint32_t out, avail; 2549 2550 if (tp->t_timers->tt_flags & TT_STOPPED) { 2551 return (1); 2552 } 2553 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 2554 /* Its not time yet */ 2555 return (0); 2556 } 2557 if (rack_progress_timeout_check(tp)) { 2558 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 2559 return (1); 2560 } 2561 /* 2562 * A TLP timer has expired. We have been idle for 2 rtts. So we now 2563 * need to figure out how to force a full MSS segment out. 2564 */ 2565 rack_log_to_event(rack, RACK_TO_FRM_TLP); 2566 counter_u64_add(rack_tlp_tot, 1); 2567 if (rack->r_state && (rack->r_state != tp->t_state)) 2568 rack_set_state(tp, rack); 2569 so = tp->t_inpcb->inp_socket; 2570 avail = sbavail(&so->so_snd); 2571 out = tp->snd_max - tp->snd_una; 2572 rack->rc_timer_up = 1; 2573 /* 2574 * If we are in recovery we can jazz out a segment if new data is 2575 * present simply by setting rc_prr_sndcnt to a segment. 2576 */ 2577 if ((avail > out) && 2578 ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { 2579 /* New data is available */ 2580 amm = avail - out; 2581 if (amm > tp->t_maxseg) { 2582 amm = tp->t_maxseg; 2583 } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 2584 /* not enough to fill a MTU and no-delay is off */ 2585 goto need_retran; 2586 } 2587 if (IN_RECOVERY(tp->t_flags)) { 2588 /* Unlikely */ 2589 old_prr_snd = rack->r_ctl.rc_prr_sndcnt; 2590 if (out + amm <= tp->snd_wnd) 2591 rack->r_ctl.rc_prr_sndcnt = amm; 2592 else 2593 goto need_retran; 2594 } else { 2595 /* Set the send-new override */ 2596 if (out + amm <= tp->snd_wnd) 2597 rack->r_ctl.rc_tlp_new_data = amm; 2598 else 2599 goto need_retran; 2600 } 2601 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2602 rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 2603 rack->r_ctl.rc_tlpsend = NULL; 2604 counter_u64_add(rack_tlp_newdata, 1); 2605 goto send; 2606 } 2607 need_retran: 2608 /* 2609 * Ok we need to arrange the last un-acked segment to be re-sent, or 2610 * optionally the first un-acked segment. 2611 */ 2612 if (rack_always_send_oldest) 2613 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 2614 else { 2615 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 2616 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { 2617 rsm = rack_find_high_nonack(rack, rsm); 2618 } 2619 } 2620 if (rsm == NULL) { 2621 counter_u64_add(rack_tlp_does_nada, 1); 2622 #ifdef TCP_BLACKBOX 2623 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); 2624 #endif 2625 goto out; 2626 } 2627 if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { 2628 /* 2629 * We need to split this the last segment in two. 2630 */ 2631 int32_t idx; 2632 struct rack_sendmap *nrsm; 2633 2634 nrsm = rack_alloc_full_limit(rack); 2635 if (nrsm == NULL) { 2636 /* 2637 * No memory to split, we will just exit and punt 2638 * off to the RXT timer. 2639 */ 2640 counter_u64_add(rack_tlp_does_nada, 1); 2641 goto out; 2642 } 2643 nrsm->r_start = (rsm->r_end - tp->t_maxseg); 2644 nrsm->r_end = rsm->r_end; 2645 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 2646 nrsm->r_flags = rsm->r_flags; 2647 nrsm->r_sndcnt = rsm->r_sndcnt; 2648 nrsm->r_rtr_bytes = 0; 2649 rsm->r_end = nrsm->r_start; 2650 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 2651 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 2652 } 2653 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 2654 if (rsm->r_in_tmap) { 2655 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 2656 nrsm->r_in_tmap = 1; 2657 } 2658 rsm->r_flags &= (~RACK_HAS_FIN); 2659 rsm = nrsm; 2660 } 2661 rack->r_ctl.rc_tlpsend = rsm; 2662 rack->r_ctl.rc_tlp_rtx_out = 1; 2663 if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { 2664 rack->r_ctl.rc_tlp_seg_send_cnt++; 2665 tp->t_rxtshift++; 2666 } else { 2667 rack->r_ctl.rc_last_tlp_seq = rsm->r_start; 2668 rack->r_ctl.rc_tlp_seg_send_cnt = 1; 2669 } 2670 send: 2671 rack->r_ctl.rc_tlp_send_cnt++; 2672 if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { 2673 /* 2674 * Can't [re]/transmit a segment we have not heard from the 2675 * peer in max times. We need the retransmit timer to take 2676 * over. 2677 */ 2678 restore: 2679 rack->r_ctl.rc_tlpsend = NULL; 2680 if (rsm) 2681 rsm->r_flags &= ~RACK_TLP; 2682 rack->r_ctl.rc_prr_sndcnt = old_prr_snd; 2683 counter_u64_add(rack_tlp_retran_fail, 1); 2684 goto out; 2685 } else if (rsm) { 2686 rsm->r_flags |= RACK_TLP; 2687 } 2688 if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && 2689 (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { 2690 /* 2691 * We don't want to send a single segment more than the max 2692 * either. 2693 */ 2694 goto restore; 2695 } 2696 rack->r_timer_override = 1; 2697 rack->r_tlp_running = 1; 2698 rack->rc_tlp_in_progress = 1; 2699 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2700 return (0); 2701 out: 2702 rack->rc_timer_up = 0; 2703 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 2704 return (0); 2705 } 2706 2707 /* 2708 * Delayed ack Timer, here we simply need to setup the 2709 * ACK_NOW flag and remove the DELACK flag. From there 2710 * the output routine will send the ack out. 2711 * 2712 * We only return 1, saying don't proceed, if all timers 2713 * are stopped (destroyed PCB?). 2714 */ 2715 static int 2716 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2717 { 2718 if (tp->t_timers->tt_flags & TT_STOPPED) { 2719 return (1); 2720 } 2721 rack_log_to_event(rack, RACK_TO_FRM_DELACK); 2722 tp->t_flags &= ~TF_DELACK; 2723 tp->t_flags |= TF_ACKNOW; 2724 TCPSTAT_INC(tcps_delack); 2725 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 2726 return (0); 2727 } 2728 2729 /* 2730 * Persists timer, here we simply need to setup the 2731 * FORCE-DATA flag the output routine will send 2732 * the one byte send. 2733 * 2734 * We only return 1, saying don't proceed, if all timers 2735 * are stopped (destroyed PCB?). 2736 */ 2737 static int 2738 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2739 { 2740 struct inpcb *inp; 2741 int32_t retval = 0; 2742 2743 inp = tp->t_inpcb; 2744 2745 if (tp->t_timers->tt_flags & TT_STOPPED) { 2746 return (1); 2747 } 2748 if (rack->rc_in_persist == 0) 2749 return (0); 2750 if (rack_progress_timeout_check(tp)) { 2751 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2752 return (1); 2753 } 2754 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 2755 /* 2756 * Persistence timer into zero window. Force a byte to be output, if 2757 * possible. 2758 */ 2759 TCPSTAT_INC(tcps_persisttimeo); 2760 /* 2761 * Hack: if the peer is dead/unreachable, we do not time out if the 2762 * window is closed. After a full backoff, drop the connection if 2763 * the idle time (no responses to probes) reaches the maximum 2764 * backoff that we would use if retransmitting. 2765 */ 2766 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 2767 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 2768 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 2769 TCPSTAT_INC(tcps_persistdrop); 2770 retval = 1; 2771 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2772 goto out; 2773 } 2774 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && 2775 tp->snd_una == tp->snd_max) 2776 rack_exit_persist(tp, rack); 2777 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 2778 /* 2779 * If the user has closed the socket then drop a persisting 2780 * connection after a much reduced timeout. 2781 */ 2782 if (tp->t_state > TCPS_CLOSE_WAIT && 2783 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 2784 retval = 1; 2785 TCPSTAT_INC(tcps_persistdrop); 2786 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2787 goto out; 2788 } 2789 tp->t_flags |= TF_FORCEDATA; 2790 out: 2791 rack_log_to_event(rack, RACK_TO_FRM_PERSIST); 2792 return (retval); 2793 } 2794 2795 /* 2796 * If a keepalive goes off, we had no other timers 2797 * happening. We always return 1 here since this 2798 * routine either drops the connection or sends 2799 * out a segment with respond. 2800 */ 2801 static int 2802 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2803 { 2804 struct tcptemp *t_template; 2805 struct inpcb *inp; 2806 2807 if (tp->t_timers->tt_flags & TT_STOPPED) { 2808 return (1); 2809 } 2810 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 2811 inp = tp->t_inpcb; 2812 rack_log_to_event(rack, RACK_TO_FRM_KEEP); 2813 /* 2814 * Keep-alive timer went off; send something or drop connection if 2815 * idle for too long. 2816 */ 2817 TCPSTAT_INC(tcps_keeptimeo); 2818 if (tp->t_state < TCPS_ESTABLISHED) 2819 goto dropit; 2820 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 2821 tp->t_state <= TCPS_CLOSING) { 2822 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 2823 goto dropit; 2824 /* 2825 * Send a packet designed to force a response if the peer is 2826 * up and reachable: either an ACK if the connection is 2827 * still alive, or an RST if the peer has closed the 2828 * connection due to timeout or reboot. Using sequence 2829 * number tp->snd_una-1 causes the transmitted zero-length 2830 * segment to lie outside the receive window; by the 2831 * protocol spec, this requires the correspondent TCP to 2832 * respond. 2833 */ 2834 TCPSTAT_INC(tcps_keepprobe); 2835 t_template = tcpip_maketemplate(inp); 2836 if (t_template) { 2837 tcp_respond(tp, t_template->tt_ipgen, 2838 &t_template->tt_t, (struct mbuf *)NULL, 2839 tp->rcv_nxt, tp->snd_una - 1, 0); 2840 free(t_template, M_TEMP); 2841 } 2842 } 2843 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 2844 return (1); 2845 dropit: 2846 TCPSTAT_INC(tcps_keepdrops); 2847 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); 2848 return (1); 2849 } 2850 2851 /* 2852 * Retransmit helper function, clear up all the ack 2853 * flags and take care of important book keeping. 2854 */ 2855 static void 2856 rack_remxt_tmr(struct tcpcb *tp) 2857 { 2858 /* 2859 * The retransmit timer went off, all sack'd blocks must be 2860 * un-acked. 2861 */ 2862 struct rack_sendmap *rsm, *trsm = NULL; 2863 struct tcp_rack *rack; 2864 int32_t cnt = 0; 2865 2866 rack = (struct tcp_rack *)tp->t_fb_ptr; 2867 rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); 2868 rack_log_to_event(rack, RACK_TO_FRM_TMR); 2869 if (rack->r_state && (rack->r_state != tp->t_state)) 2870 rack_set_state(tp, rack); 2871 /* 2872 * Ideally we would like to be able to 2873 * mark SACK-PASS on anything not acked here. 2874 * However, if we do that we would burst out 2875 * all that data 1ms apart. This would be unwise, 2876 * so for now we will just let the normal rxt timer 2877 * and tlp timer take care of it. 2878 */ 2879 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 2880 if (rsm->r_flags & RACK_ACKED) { 2881 cnt++; 2882 rsm->r_sndcnt = 0; 2883 if (rsm->r_in_tmap == 0) { 2884 /* We must re-add it back to the tlist */ 2885 if (trsm == NULL) { 2886 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 2887 } else { 2888 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); 2889 } 2890 rsm->r_in_tmap = 1; 2891 trsm = rsm; 2892 } 2893 } 2894 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); 2895 } 2896 /* Clear the count (we just un-acked them) */ 2897 rack->r_ctl.rc_sacked = 0; 2898 /* Clear the tlp rtx mark */ 2899 rack->r_ctl.rc_tlp_rtx_out = 0; 2900 rack->r_ctl.rc_tlp_seg_send_cnt = 0; 2901 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); 2902 /* Setup so we send one segment */ 2903 if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) 2904 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 2905 rack->r_timer_override = 1; 2906 } 2907 2908 /* 2909 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 2910 * we will setup to retransmit the lowest seq number outstanding. 2911 */ 2912 static int 2913 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) 2914 { 2915 int32_t rexmt; 2916 struct inpcb *inp; 2917 int32_t retval = 0; 2918 2919 inp = tp->t_inpcb; 2920 if (tp->t_timers->tt_flags & TT_STOPPED) { 2921 return (1); 2922 } 2923 if (rack_progress_timeout_check(tp)) { 2924 tcp_set_inp_to_drop(inp, ETIMEDOUT); 2925 return (1); 2926 } 2927 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 2928 if (TCPS_HAVEESTABLISHED(tp->t_state) && 2929 (tp->snd_una == tp->snd_max)) { 2930 /* Nothing outstanding .. nothing to do */ 2931 return (0); 2932 } 2933 /* 2934 * Retransmission timer went off. Message has not been acked within 2935 * retransmit interval. Back off to a longer retransmit interval 2936 * and retransmit one segment. 2937 */ 2938 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { 2939 tp->t_rxtshift = TCP_MAXRXTSHIFT; 2940 TCPSTAT_INC(tcps_timeoutdrop); 2941 retval = 1; 2942 tcp_set_inp_to_drop(rack->rc_inp, 2943 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); 2944 goto out; 2945 } 2946 rack_remxt_tmr(tp); 2947 if (tp->t_state == TCPS_SYN_SENT) { 2948 /* 2949 * If the SYN was retransmitted, indicate CWND to be limited 2950 * to 1 segment in cc_conn_init(). 2951 */ 2952 tp->snd_cwnd = 1; 2953 } else if (tp->t_rxtshift == 1) { 2954 /* 2955 * first retransmit; record ssthresh and cwnd so they can be 2956 * recovered if this turns out to be a "bad" retransmit. A 2957 * retransmit is considered "bad" if an ACK for this segment 2958 * is received within RTT/2 interval; the assumption here is 2959 * that the ACK was already in flight. See "On Estimating 2960 * End-to-End Network Path Properties" by Allman and Paxson 2961 * for more details. 2962 */ 2963 tp->snd_cwnd_prev = tp->snd_cwnd; 2964 tp->snd_ssthresh_prev = tp->snd_ssthresh; 2965 tp->snd_recover_prev = tp->snd_recover; 2966 if (IN_FASTRECOVERY(tp->t_flags)) 2967 tp->t_flags |= TF_WASFRECOVERY; 2968 else 2969 tp->t_flags &= ~TF_WASFRECOVERY; 2970 if (IN_CONGRECOVERY(tp->t_flags)) 2971 tp->t_flags |= TF_WASCRECOVERY; 2972 else 2973 tp->t_flags &= ~TF_WASCRECOVERY; 2974 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 2975 tp->t_flags |= TF_PREVVALID; 2976 } else 2977 tp->t_flags &= ~TF_PREVVALID; 2978 TCPSTAT_INC(tcps_rexmttimeo); 2979 if ((tp->t_state == TCPS_SYN_SENT) || 2980 (tp->t_state == TCPS_SYN_RECEIVED)) 2981 rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]); 2982 else 2983 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 2984 TCPT_RANGESET(tp->t_rxtcur, rexmt, 2985 max(MSEC_2_TICKS(rack_rto_min), rexmt), 2986 MSEC_2_TICKS(rack_rto_max)); 2987 /* 2988 * We enter the path for PLMTUD if connection is established or, if 2989 * connection is FIN_WAIT_1 status, reason for the last is that if 2990 * amount of data we send is very small, we could send it in couple 2991 * of packets and process straight to FIN. In that case we won't 2992 * catch ESTABLISHED state. 2993 */ 2994 if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) 2995 || (tp->t_state == TCPS_FIN_WAIT_1))) { 2996 #ifdef INET6 2997 int32_t isipv6; 2998 #endif 2999 3000 /* 3001 * Idea here is that at each stage of mtu probe (usually, 3002 * 1448 -> 1188 -> 524) should be given 2 chances to recover 3003 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 3004 * should take care of that. 3005 */ 3006 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 3007 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 3008 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 3009 tp->t_rxtshift % 2 == 0)) { 3010 /* 3011 * Enter Path MTU Black-hole Detection mechanism: - 3012 * Disable Path MTU Discovery (IP "DF" bit). - 3013 * Reduce MTU to lower value than what we negotiated 3014 * with peer. 3015 */ 3016 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 3017 /* Record that we may have found a black hole. */ 3018 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 3019 /* Keep track of previous MSS. */ 3020 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 3021 } 3022 3023 /* 3024 * Reduce the MSS to blackhole value or to the 3025 * default in an attempt to retransmit. 3026 */ 3027 #ifdef INET6 3028 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; 3029 if (isipv6 && 3030 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 3031 /* Use the sysctl tuneable blackhole MSS. */ 3032 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 3033 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 3034 } else if (isipv6) { 3035 /* Use the default MSS. */ 3036 tp->t_maxseg = V_tcp_v6mssdflt; 3037 /* 3038 * Disable Path MTU Discovery when we switch 3039 * to minmss. 3040 */ 3041 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3042 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3043 } 3044 #endif 3045 #if defined(INET6) && defined(INET) 3046 else 3047 #endif 3048 #ifdef INET 3049 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 3050 /* Use the sysctl tuneable blackhole MSS. */ 3051 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 3052 TCPSTAT_INC(tcps_pmtud_blackhole_activated); 3053 } else { 3054 /* Use the default MSS. */ 3055 tp->t_maxseg = V_tcp_mssdflt; 3056 /* 3057 * Disable Path MTU Discovery when we switch 3058 * to minmss. 3059 */ 3060 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 3061 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 3062 } 3063 #endif 3064 } else { 3065 /* 3066 * If further retransmissions are still unsuccessful 3067 * with a lowered MTU, maybe this isn't a blackhole 3068 * and we restore the previous MSS and blackhole 3069 * detection flags. The limit '6' is determined by 3070 * giving each probe stage (1448, 1188, 524) 2 3071 * chances to recover. 3072 */ 3073 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 3074 (tp->t_rxtshift >= 6)) { 3075 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 3076 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 3077 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 3078 TCPSTAT_INC(tcps_pmtud_blackhole_failed); 3079 } 3080 } 3081 } 3082 /* 3083 * Disable RFC1323 and SACK if we haven't got any response to our 3084 * third SYN to work-around some broken terminal servers (most of 3085 * which have hopefully been retired) that have bad VJ header 3086 * compression code which trashes TCP segments containing 3087 * unknown-to-them TCP options. 3088 */ 3089 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 3090 (tp->t_rxtshift == 3)) 3091 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 3092 /* 3093 * If we backed off this far, our srtt estimate is probably bogus. 3094 * Clobber it so we'll take the next rtt measurement as our srtt; 3095 * move the current srtt into rttvar to keep the current retransmit 3096 * times until then. 3097 */ 3098 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 3099 #ifdef INET6 3100 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 3101 in6_losing(tp->t_inpcb); 3102 else 3103 #endif 3104 in_losing(tp->t_inpcb); 3105 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 3106 tp->t_srtt = 0; 3107 } 3108 if (rack_use_sack_filter) 3109 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 3110 tp->snd_recover = tp->snd_max; 3111 tp->t_flags |= TF_ACKNOW; 3112 tp->t_rtttime = 0; 3113 rack_cong_signal(tp, NULL, CC_RTO); 3114 out: 3115 return (retval); 3116 } 3117 3118 static int 3119 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) 3120 { 3121 int32_t ret = 0; 3122 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 3123 3124 if (timers == 0) { 3125 return (0); 3126 } 3127 if (tp->t_state == TCPS_LISTEN) { 3128 /* no timers on listen sockets */ 3129 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 3130 return (0); 3131 return (1); 3132 } 3133 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { 3134 uint32_t left; 3135 3136 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 3137 ret = -1; 3138 rack_log_to_processing(rack, cts, ret, 0); 3139 return (0); 3140 } 3141 if (hpts_calling == 0) { 3142 ret = -2; 3143 rack_log_to_processing(rack, cts, ret, 0); 3144 return (0); 3145 } 3146 /* 3147 * Ok our timer went off early and we are not paced false 3148 * alarm, go back to sleep. 3149 */ 3150 ret = -3; 3151 left = rack->r_ctl.rc_timer_exp - cts; 3152 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); 3153 rack_log_to_processing(rack, cts, ret, left); 3154 rack->rc_last_pto_set = 0; 3155 return (1); 3156 } 3157 rack->rc_tmr_stopped = 0; 3158 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 3159 if (timers & PACE_TMR_DELACK) { 3160 ret = rack_timeout_delack(tp, rack, cts); 3161 } else if (timers & PACE_TMR_RACK) { 3162 ret = rack_timeout_rack(tp, rack, cts); 3163 } else if (timers & PACE_TMR_TLP) { 3164 ret = rack_timeout_tlp(tp, rack, cts); 3165 } else if (timers & PACE_TMR_RXT) { 3166 ret = rack_timeout_rxt(tp, rack, cts); 3167 } else if (timers & PACE_TMR_PERSIT) { 3168 ret = rack_timeout_persist(tp, rack, cts); 3169 } else if (timers & PACE_TMR_KEEP) { 3170 ret = rack_timeout_keepalive(tp, rack, cts); 3171 } 3172 rack_log_to_processing(rack, cts, ret, timers); 3173 return (ret); 3174 } 3175 3176 static void 3177 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) 3178 { 3179 uint8_t hpts_removed = 0; 3180 3181 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 3182 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { 3183 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3184 hpts_removed = 1; 3185 } 3186 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 3187 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 3188 if (rack->rc_inp->inp_in_hpts && 3189 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { 3190 /* 3191 * Canceling timer's when we have no output being 3192 * paced. We also must remove ourselves from the 3193 * hpts. 3194 */ 3195 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 3196 hpts_removed = 1; 3197 } 3198 rack_log_to_cancel(rack, hpts_removed, line); 3199 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 3200 } 3201 } 3202 3203 static void 3204 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) 3205 { 3206 return; 3207 } 3208 3209 static int 3210 rack_stopall(struct tcpcb *tp) 3211 { 3212 struct tcp_rack *rack; 3213 rack = (struct tcp_rack *)tp->t_fb_ptr; 3214 rack->t_timers_stopped = 1; 3215 return (0); 3216 } 3217 3218 static void 3219 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 3220 { 3221 return; 3222 } 3223 3224 static int 3225 rack_timer_active(struct tcpcb *tp, uint32_t timer_type) 3226 { 3227 return (0); 3228 } 3229 3230 static void 3231 rack_stop_all_timers(struct tcpcb *tp) 3232 { 3233 struct tcp_rack *rack; 3234 3235 /* 3236 * Assure no timers are running. 3237 */ 3238 if (tcp_timer_active(tp, TT_PERSIST)) { 3239 /* We enter in persists, set the flag appropriately */ 3240 rack = (struct tcp_rack *)tp->t_fb_ptr; 3241 rack->rc_in_persist = 1; 3242 } 3243 tcp_timer_suspend(tp, TT_PERSIST); 3244 tcp_timer_suspend(tp, TT_REXMT); 3245 tcp_timer_suspend(tp, TT_KEEP); 3246 tcp_timer_suspend(tp, TT_DELACK); 3247 } 3248 3249 static void 3250 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, 3251 struct rack_sendmap *rsm, uint32_t ts) 3252 { 3253 int32_t idx; 3254 3255 rsm->r_rtr_cnt++; 3256 rsm->r_sndcnt++; 3257 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { 3258 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; 3259 rsm->r_flags |= RACK_OVERMAX; 3260 } 3261 if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { 3262 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 3263 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 3264 } 3265 idx = rsm->r_rtr_cnt - 1; 3266 rsm->r_tim_lastsent[idx] = ts; 3267 if (rsm->r_flags & RACK_ACKED) { 3268 /* Problably MTU discovery messing with us */ 3269 rsm->r_flags &= ~RACK_ACKED; 3270 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 3271 } 3272 if (rsm->r_in_tmap) { 3273 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3274 } 3275 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3276 rsm->r_in_tmap = 1; 3277 if (rsm->r_flags & RACK_SACK_PASSED) { 3278 /* We have retransmitted due to the SACK pass */ 3279 rsm->r_flags &= ~RACK_SACK_PASSED; 3280 rsm->r_flags |= RACK_WAS_SACKPASS; 3281 } 3282 /* Update memory for next rtr */ 3283 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3284 } 3285 3286 3287 static uint32_t 3288 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, 3289 struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) 3290 { 3291 /* 3292 * We (re-)transmitted starting at rsm->r_start for some length 3293 * (possibly less than r_end. 3294 */ 3295 struct rack_sendmap *nrsm; 3296 uint32_t c_end; 3297 int32_t len; 3298 int32_t idx; 3299 3300 len = *lenp; 3301 c_end = rsm->r_start + len; 3302 if (SEQ_GEQ(c_end, rsm->r_end)) { 3303 /* 3304 * We retransmitted the whole piece or more than the whole 3305 * slopping into the next rsm. 3306 */ 3307 rack_update_rsm(tp, rack, rsm, ts); 3308 if (c_end == rsm->r_end) { 3309 *lenp = 0; 3310 return (0); 3311 } else { 3312 int32_t act_len; 3313 3314 /* Hangs over the end return whats left */ 3315 act_len = rsm->r_end - rsm->r_start; 3316 *lenp = (len - act_len); 3317 return (rsm->r_end); 3318 } 3319 /* We don't get out of this block. */ 3320 } 3321 /* 3322 * Here we retransmitted less than the whole thing which means we 3323 * have to split this into what was transmitted and what was not. 3324 */ 3325 nrsm = rack_alloc_full_limit(rack); 3326 if (nrsm == NULL) { 3327 /* 3328 * We can't get memory, so lets not proceed. 3329 */ 3330 *lenp = 0; 3331 return (0); 3332 } 3333 /* 3334 * So here we are going to take the original rsm and make it what we 3335 * retransmitted. nrsm will be the tail portion we did not 3336 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 3337 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 3338 * 1, 6 and the new piece will be 6, 11. 3339 */ 3340 nrsm->r_start = c_end; 3341 nrsm->r_end = rsm->r_end; 3342 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3343 nrsm->r_flags = rsm->r_flags; 3344 nrsm->r_sndcnt = rsm->r_sndcnt; 3345 nrsm->r_rtr_bytes = 0; 3346 rsm->r_end = c_end; 3347 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3348 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3349 } 3350 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3351 if (rsm->r_in_tmap) { 3352 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3353 nrsm->r_in_tmap = 1; 3354 } 3355 rsm->r_flags &= (~RACK_HAS_FIN); 3356 rack_update_rsm(tp, rack, rsm, ts); 3357 *lenp = 0; 3358 return (0); 3359 } 3360 3361 3362 static void 3363 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, 3364 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, 3365 uint8_t pass, struct rack_sendmap *hintrsm) 3366 { 3367 struct tcp_rack *rack; 3368 struct rack_sendmap *rsm, *nrsm; 3369 register uint32_t snd_max, snd_una; 3370 int32_t idx; 3371 3372 /* 3373 * Add to the RACK log of packets in flight or retransmitted. If 3374 * there is a TS option we will use the TS echoed, if not we will 3375 * grab a TS. 3376 * 3377 * Retransmissions will increment the count and move the ts to its 3378 * proper place. Note that if options do not include TS's then we 3379 * won't be able to effectively use the ACK for an RTT on a retran. 3380 * 3381 * Notes about r_start and r_end. Lets consider a send starting at 3382 * sequence 1 for 10 bytes. In such an example the r_start would be 3383 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 3384 * This means that r_end is actually the first sequence for the next 3385 * slot (11). 3386 * 3387 */ 3388 /* 3389 * If err is set what do we do XXXrrs? should we not add the thing? 3390 * -- i.e. return if err != 0 or should we pretend we sent it? -- 3391 * i.e. proceed with add ** do this for now. 3392 */ 3393 INP_WLOCK_ASSERT(tp->t_inpcb); 3394 if (err) 3395 /* 3396 * We don't log errors -- we could but snd_max does not 3397 * advance in this case either. 3398 */ 3399 return; 3400 3401 if (th_flags & TH_RST) { 3402 /* 3403 * We don't log resets and we return immediately from 3404 * sending 3405 */ 3406 return; 3407 } 3408 rack = (struct tcp_rack *)tp->t_fb_ptr; 3409 snd_una = tp->snd_una; 3410 if (SEQ_LEQ((seq_out + len), snd_una)) { 3411 /* Are sending an old segment to induce an ack (keep-alive)? */ 3412 return; 3413 } 3414 if (SEQ_LT(seq_out, snd_una)) { 3415 /* huh? should we panic? */ 3416 uint32_t end; 3417 3418 end = seq_out + len; 3419 seq_out = snd_una; 3420 len = end - seq_out; 3421 } 3422 snd_max = tp->snd_max; 3423 if (th_flags & (TH_SYN | TH_FIN)) { 3424 /* 3425 * The call to rack_log_output is made before bumping 3426 * snd_max. This means we can record one extra byte on a SYN 3427 * or FIN if seq_out is adding more on and a FIN is present 3428 * (and we are not resending). 3429 */ 3430 if (th_flags & TH_SYN) 3431 len++; 3432 if (th_flags & TH_FIN) 3433 len++; 3434 if (SEQ_LT(snd_max, tp->snd_nxt)) { 3435 /* 3436 * The add/update as not been done for the FIN/SYN 3437 * yet. 3438 */ 3439 snd_max = tp->snd_nxt; 3440 } 3441 } 3442 if (len == 0) { 3443 /* We don't log zero window probes */ 3444 return; 3445 } 3446 rack->r_ctl.rc_time_last_sent = ts; 3447 if (IN_RECOVERY(tp->t_flags)) { 3448 rack->r_ctl.rc_prr_out += len; 3449 } 3450 /* First question is it a retransmission? */ 3451 if (seq_out == snd_max) { 3452 again: 3453 rsm = rack_alloc(rack); 3454 if (rsm == NULL) { 3455 /* 3456 * Hmm out of memory and the tcb got destroyed while 3457 * we tried to wait. 3458 */ 3459 return; 3460 } 3461 if (th_flags & TH_FIN) { 3462 rsm->r_flags = RACK_HAS_FIN; 3463 } else { 3464 rsm->r_flags = 0; 3465 } 3466 rsm->r_tim_lastsent[0] = ts; 3467 rsm->r_rtr_cnt = 1; 3468 rsm->r_rtr_bytes = 0; 3469 rsm->r_start = seq_out; 3470 rsm->r_end = rsm->r_start + len; 3471 rsm->r_sndcnt = 0; 3472 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 3473 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 3474 rsm->r_in_tmap = 1; 3475 return; 3476 } 3477 /* 3478 * If we reach here its a retransmission and we need to find it. 3479 */ 3480 more: 3481 if (hintrsm && (hintrsm->r_start == seq_out)) { 3482 rsm = hintrsm; 3483 hintrsm = NULL; 3484 } else if (rack->r_ctl.rc_next) { 3485 /* We have a hint from a previous run */ 3486 rsm = rack->r_ctl.rc_next; 3487 } else { 3488 /* No hints sorry */ 3489 rsm = NULL; 3490 } 3491 if ((rsm) && (rsm->r_start == seq_out)) { 3492 /* 3493 * We used rc_next or hintrsm to retransmit, hopefully the 3494 * likely case. 3495 */ 3496 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3497 if (len == 0) { 3498 return; 3499 } else { 3500 goto more; 3501 } 3502 } 3503 /* Ok it was not the last pointer go through it the hard way. */ 3504 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3505 if (rsm->r_start == seq_out) { 3506 seq_out = rack_update_entry(tp, rack, rsm, ts, &len); 3507 rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 3508 if (len == 0) { 3509 return; 3510 } else { 3511 continue; 3512 } 3513 } 3514 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 3515 /* Transmitted within this piece */ 3516 /* 3517 * Ok we must split off the front and then let the 3518 * update do the rest 3519 */ 3520 nrsm = rack_alloc_full_limit(rack); 3521 if (nrsm == NULL) { 3522 rack_update_rsm(tp, rack, rsm, ts); 3523 return; 3524 } 3525 /* 3526 * copy rsm to nrsm and then trim the front of rsm 3527 * to not include this part. 3528 */ 3529 nrsm->r_start = seq_out; 3530 nrsm->r_end = rsm->r_end; 3531 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 3532 nrsm->r_flags = rsm->r_flags; 3533 nrsm->r_sndcnt = rsm->r_sndcnt; 3534 nrsm->r_rtr_bytes = 0; 3535 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 3536 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 3537 } 3538 rsm->r_end = nrsm->r_start; 3539 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 3540 if (rsm->r_in_tmap) { 3541 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 3542 nrsm->r_in_tmap = 1; 3543 } 3544 rsm->r_flags &= (~RACK_HAS_FIN); 3545 seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); 3546 if (len == 0) { 3547 return; 3548 } 3549 } 3550 } 3551 /* 3552 * Hmm not found in map did they retransmit both old and on into the 3553 * new? 3554 */ 3555 if (seq_out == tp->snd_max) { 3556 goto again; 3557 } else if (SEQ_LT(seq_out, tp->snd_max)) { 3558 #ifdef INVARIANTS 3559 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 3560 seq_out, len, tp->snd_una, tp->snd_max); 3561 printf("Starting Dump of all rack entries\n"); 3562 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { 3563 printf("rsm:%p start:%u end:%u\n", 3564 rsm, rsm->r_start, rsm->r_end); 3565 } 3566 printf("Dump complete\n"); 3567 panic("seq_out not found rack:%p tp:%p", 3568 rack, tp); 3569 #endif 3570 } else { 3571 #ifdef INVARIANTS 3572 /* 3573 * Hmm beyond sndmax? (only if we are using the new rtt-pack 3574 * flag) 3575 */ 3576 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 3577 seq_out, len, tp->snd_max, tp); 3578 #endif 3579 } 3580 } 3581 3582 /* 3583 * Record one of the RTT updates from an ack into 3584 * our sample structure. 3585 */ 3586 static void 3587 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) 3588 { 3589 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3590 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { 3591 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; 3592 } 3593 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || 3594 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { 3595 rack->r_ctl.rack_rs.rs_rtt_highest = rtt; 3596 } 3597 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; 3598 rack->r_ctl.rack_rs.rs_rtt_tot += rtt; 3599 rack->r_ctl.rack_rs.rs_rtt_cnt++; 3600 } 3601 3602 /* 3603 * Collect new round-trip time estimate 3604 * and update averages and current timeout. 3605 */ 3606 static void 3607 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) 3608 { 3609 int32_t delta; 3610 uint32_t o_srtt, o_var; 3611 int32_t rtt; 3612 3613 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) 3614 /* No valid sample */ 3615 return; 3616 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { 3617 /* We are to use the lowest RTT seen in a single ack */ 3618 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; 3619 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { 3620 /* We are to use the highest RTT seen in a single ack */ 3621 rtt = rack->r_ctl.rack_rs.rs_rtt_highest; 3622 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { 3623 /* We are to use the average RTT seen in a single ack */ 3624 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / 3625 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); 3626 } else { 3627 #ifdef INVARIANTS 3628 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); 3629 #endif 3630 return; 3631 } 3632 if (rtt == 0) 3633 rtt = 1; 3634 rack_log_rtt_sample(rack, rtt); 3635 o_srtt = tp->t_srtt; 3636 o_var = tp->t_rttvar; 3637 rack = (struct tcp_rack *)tp->t_fb_ptr; 3638 if (tp->t_srtt != 0) { 3639 /* 3640 * srtt is stored as fixed point with 5 bits after the 3641 * binary point (i.e., scaled by 8). The following magic is 3642 * equivalent to the smoothing algorithm in rfc793 with an 3643 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 3644 * Adjust rtt to origin 0. 3645 */ 3646 delta = ((rtt - 1) << TCP_DELTA_SHIFT) 3647 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 3648 3649 tp->t_srtt += delta; 3650 if (tp->t_srtt <= 0) 3651 tp->t_srtt = 1; 3652 3653 /* 3654 * We accumulate a smoothed rtt variance (actually, a 3655 * smoothed mean difference), then set the retransmit timer 3656 * to smoothed rtt + 4 times the smoothed variance. rttvar 3657 * is stored as fixed point with 4 bits after the binary 3658 * point (scaled by 16). The following is equivalent to 3659 * rfc793 smoothing with an alpha of .75 (rttvar = 3660 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 3661 * wired-in beta. 3662 */ 3663 if (delta < 0) 3664 delta = -delta; 3665 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 3666 tp->t_rttvar += delta; 3667 if (tp->t_rttvar <= 0) 3668 tp->t_rttvar = 1; 3669 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 3670 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3671 } else { 3672 /* 3673 * No rtt measurement yet - use the unsmoothed rtt. Set the 3674 * variance to half the rtt (so our first retransmit happens 3675 * at 3*rtt). 3676 */ 3677 tp->t_srtt = rtt << TCP_RTT_SHIFT; 3678 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); 3679 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 3680 } 3681 TCPSTAT_INC(tcps_rttupdated); 3682 rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); 3683 tp->t_rttupdated++; 3684 #ifdef NETFLIX_STATS 3685 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); 3686 #endif 3687 tp->t_rxtshift = 0; 3688 3689 /* 3690 * the retransmit should happen at rtt + 4 * rttvar. Because of the 3691 * way we do the smoothing, srtt and rttvar will each average +1/2 3692 * tick of bias. When we compute the retransmit timer, we want 1/2 3693 * tick of rounding and 1 extra tick because of +-1/2 tick 3694 * uncertainty in the firing of the timer. The bias will give us 3695 * exactly the 1.5 tick we need. But, because the bias is 3696 * statistical, we have to test that we don't drop below the minimum 3697 * feasible timer (which is 2 ticks). 3698 */ 3699 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 3700 max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); 3701 tp->t_softerror = 0; 3702 } 3703 3704 static void 3705 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, 3706 uint32_t t, uint32_t cts) 3707 { 3708 /* 3709 * For this RSM, we acknowledged the data from a previous 3710 * transmission, not the last one we made. This means we did a false 3711 * retransmit. 3712 */ 3713 struct tcp_rack *rack; 3714 3715 if (rsm->r_flags & RACK_HAS_FIN) { 3716 /* 3717 * The sending of the FIN often is multiple sent when we 3718 * have everything outstanding ack'd. We ignore this case 3719 * since its over now. 3720 */ 3721 return; 3722 } 3723 if (rsm->r_flags & RACK_TLP) { 3724 /* 3725 * We expect TLP's to have this occur. 3726 */ 3727 return; 3728 } 3729 rack = (struct tcp_rack *)tp->t_fb_ptr; 3730 /* should we undo cc changes and exit recovery? */ 3731 if (IN_RECOVERY(tp->t_flags)) { 3732 if (rack->r_ctl.rc_rsm_start == rsm->r_start) { 3733 /* 3734 * Undo what we ratched down and exit recovery if 3735 * possible 3736 */ 3737 EXIT_RECOVERY(tp->t_flags); 3738 tp->snd_recover = tp->snd_una; 3739 if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) 3740 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; 3741 if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) 3742 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; 3743 } 3744 } 3745 if (rsm->r_flags & RACK_WAS_SACKPASS) { 3746 /* 3747 * We retransmitted based on a sack and the earlier 3748 * retransmission ack'd it - re-ordering is occuring. 3749 */ 3750 counter_u64_add(rack_reorder_seen, 1); 3751 rack->r_ctl.rc_reorder_ts = cts; 3752 } 3753 counter_u64_add(rack_badfr, 1); 3754 counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); 3755 } 3756 3757 3758 static int 3759 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, 3760 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) 3761 { 3762 int32_t i; 3763 uint32_t t; 3764 3765 if (rsm->r_flags & RACK_ACKED) 3766 /* Already done */ 3767 return (0); 3768 3769 3770 if ((rsm->r_rtr_cnt == 1) || 3771 ((ack_type == CUM_ACKED) && 3772 (to->to_flags & TOF_TS) && 3773 (to->to_tsecr) && 3774 (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) 3775 ) { 3776 /* 3777 * We will only find a matching timestamp if its cum-acked. 3778 * But if its only one retransmission its for-sure matching 3779 * :-) 3780 */ 3781 t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3782 if ((int)t <= 0) 3783 t = 1; 3784 if (!tp->t_rttlow || tp->t_rttlow > t) 3785 tp->t_rttlow = t; 3786 if (!rack->r_ctl.rc_rack_min_rtt || 3787 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3788 rack->r_ctl.rc_rack_min_rtt = t; 3789 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3790 rack->r_ctl.rc_rack_min_rtt = 1; 3791 } 3792 } 3793 tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); 3794 if ((rsm->r_flags & RACK_TLP) && 3795 (!IN_RECOVERY(tp->t_flags))) { 3796 /* Segment was a TLP and our retrans matched */ 3797 if (rack->r_ctl.rc_tlp_cwnd_reduce) { 3798 rack->r_ctl.rc_rsm_start = tp->snd_max; 3799 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 3800 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 3801 rack_cong_signal(tp, NULL, CC_NDUPACK); 3802 /* 3803 * When we enter recovery we need to assure 3804 * we send one packet. 3805 */ 3806 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 3807 } else 3808 rack->r_ctl.rc_tlp_rtx_out = 0; 3809 } 3810 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3811 /* New more recent rack_tmit_time */ 3812 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3813 rack->rc_rack_rtt = t; 3814 } 3815 return (1); 3816 } 3817 /* 3818 * We clear the soft/rxtshift since we got an ack. 3819 * There is no assurance we will call the commit() function 3820 * so we need to clear these to avoid incorrect handling. 3821 */ 3822 tp->t_rxtshift = 0; 3823 tp->t_softerror = 0; 3824 if ((to->to_flags & TOF_TS) && 3825 (ack_type == CUM_ACKED) && 3826 (to->to_tsecr) && 3827 ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { 3828 /* 3829 * Now which timestamp does it match? In this block the ACK 3830 * must be coming from a previous transmission. 3831 */ 3832 for (i = 0; i < rsm->r_rtr_cnt; i++) { 3833 if (rsm->r_tim_lastsent[i] == to->to_tsecr) { 3834 t = cts - rsm->r_tim_lastsent[i]; 3835 if ((int)t <= 0) 3836 t = 1; 3837 if ((i + 1) < rsm->r_rtr_cnt) { 3838 /* Likely */ 3839 rack_earlier_retran(tp, rsm, t, cts); 3840 } 3841 if (!tp->t_rttlow || tp->t_rttlow > t) 3842 tp->t_rttlow = t; 3843 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3844 rack->r_ctl.rc_rack_min_rtt = t; 3845 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3846 rack->r_ctl.rc_rack_min_rtt = 1; 3847 } 3848 } 3849 /* 3850 * Note the following calls to 3851 * tcp_rack_xmit_timer() are being commented 3852 * out for now. They give us no more accuracy 3853 * and often lead to a wrong choice. We have 3854 * enough samples that have not been 3855 * retransmitted. I leave the commented out 3856 * code in here in case in the future we 3857 * decide to add it back (though I can't forsee 3858 * doing that). That way we will easily see 3859 * where they need to be placed. 3860 */ 3861 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, 3862 rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { 3863 /* New more recent rack_tmit_time */ 3864 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 3865 rack->rc_rack_rtt = t; 3866 } 3867 return (1); 3868 } 3869 } 3870 goto ts_not_found; 3871 } else { 3872 /* 3873 * Ok its a SACK block that we retransmitted. or a windows 3874 * machine without timestamps. We can tell nothing from the 3875 * time-stamp since its not there or the time the peer last 3876 * recieved a segment that moved forward its cum-ack point. 3877 */ 3878 ts_not_found: 3879 i = rsm->r_rtr_cnt - 1; 3880 t = cts - rsm->r_tim_lastsent[i]; 3881 if ((int)t <= 0) 3882 t = 1; 3883 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3884 /* 3885 * We retransmitted and the ack came back in less 3886 * than the smallest rtt we have observed. We most 3887 * likey did an improper retransmit as outlined in 3888 * 4.2 Step 3 point 2 in the rack-draft. 3889 */ 3890 i = rsm->r_rtr_cnt - 2; 3891 t = cts - rsm->r_tim_lastsent[i]; 3892 rack_earlier_retran(tp, rsm, t, cts); 3893 } else if (rack->r_ctl.rc_rack_min_rtt) { 3894 /* 3895 * We retransmitted it and the retransmit did the 3896 * job. 3897 */ 3898 if (!rack->r_ctl.rc_rack_min_rtt || 3899 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { 3900 rack->r_ctl.rc_rack_min_rtt = t; 3901 if (rack->r_ctl.rc_rack_min_rtt == 0) { 3902 rack->r_ctl.rc_rack_min_rtt = 1; 3903 } 3904 } 3905 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { 3906 /* New more recent rack_tmit_time */ 3907 rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; 3908 rack->rc_rack_rtt = t; 3909 } 3910 return (1); 3911 } 3912 } 3913 return (0); 3914 } 3915 3916 /* 3917 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 3918 */ 3919 static void 3920 rack_log_sack_passed(struct tcpcb *tp, 3921 struct tcp_rack *rack, struct rack_sendmap *rsm) 3922 { 3923 struct rack_sendmap *nrsm; 3924 uint32_t ts; 3925 int32_t idx; 3926 3927 idx = rsm->r_rtr_cnt - 1; 3928 ts = rsm->r_tim_lastsent[idx]; 3929 nrsm = rsm; 3930 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, 3931 rack_head, r_tnext) { 3932 if (nrsm == rsm) { 3933 /* Skip orginal segment he is acked */ 3934 continue; 3935 } 3936 if (nrsm->r_flags & RACK_ACKED) { 3937 /* Skip ack'd segments */ 3938 continue; 3939 } 3940 if (nrsm->r_flags & RACK_SACK_PASSED) { 3941 /* 3942 * We found one that is already marked 3943 * passed, we have been here before and 3944 * so all others below this are marked. 3945 */ 3946 break; 3947 } 3948 idx = nrsm->r_rtr_cnt - 1; 3949 if (ts == nrsm->r_tim_lastsent[idx]) { 3950 /* 3951 * For this case lets use seq no, if we sent in a 3952 * big block (TSO) we would have a bunch of segments 3953 * sent at the same time. 3954 * 3955 * We would only get a report if its SEQ is earlier. 3956 * If we have done multiple retransmits the times 3957 * would not be equal. 3958 */ 3959 if (SEQ_LT(nrsm->r_start, rsm->r_start)) { 3960 nrsm->r_flags |= RACK_SACK_PASSED; 3961 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3962 } 3963 } else { 3964 /* 3965 * Here they were sent at different times, not a big 3966 * block. Since we transmitted this one later and 3967 * see it sack'd then this must also be missing (or 3968 * we would have gotten a sack block for it) 3969 */ 3970 nrsm->r_flags |= RACK_SACK_PASSED; 3971 nrsm->r_flags &= ~RACK_WAS_SACKPASS; 3972 } 3973 } 3974 } 3975 3976 static uint32_t 3977 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, 3978 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) 3979 { 3980 int32_t idx; 3981 int32_t times = 0; 3982 uint32_t start, end, changed = 0; 3983 struct rack_sendmap *rsm, *nrsm; 3984 int32_t used_ref = 1; 3985 3986 start = sack->start; 3987 end = sack->end; 3988 rsm = *prsm; 3989 if (rsm && SEQ_LT(start, rsm->r_start)) { 3990 TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { 3991 if (SEQ_GEQ(start, rsm->r_start) && 3992 SEQ_LT(start, rsm->r_end)) { 3993 goto do_rest_ofb; 3994 } 3995 } 3996 } 3997 if (rsm == NULL) { 3998 start_at_beginning: 3999 rsm = NULL; 4000 used_ref = 0; 4001 } 4002 /* First lets locate the block where this guy is */ 4003 TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { 4004 if (SEQ_GEQ(start, rsm->r_start) && 4005 SEQ_LT(start, rsm->r_end)) { 4006 break; 4007 } 4008 } 4009 do_rest_ofb: 4010 if (rsm == NULL) { 4011 /* 4012 * This happens when we get duplicate sack blocks with the 4013 * same end. For example SACK 4: 100 SACK 3: 100 The sort 4014 * will not change there location so we would just start at 4015 * the end of the first one and get lost. 4016 */ 4017 if (tp->t_flags & TF_SENTFIN) { 4018 /* 4019 * Check to see if we have not logged the FIN that 4020 * went out. 4021 */ 4022 nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4023 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 4024 /* 4025 * Ok we did not get the FIN logged. 4026 */ 4027 nrsm->r_end++; 4028 rsm = nrsm; 4029 goto do_rest_ofb; 4030 } 4031 } 4032 if (times == 1) { 4033 #ifdef INVARIANTS 4034 panic("tp:%p rack:%p sack:%p to:%p prsm:%p", 4035 tp, rack, sack, to, prsm); 4036 #else 4037 goto out; 4038 #endif 4039 } 4040 times++; 4041 counter_u64_add(rack_sack_proc_restart, 1); 4042 goto start_at_beginning; 4043 } 4044 /* Ok we have an ACK for some piece of rsm */ 4045 if (rsm->r_start != start) { 4046 /* 4047 * Need to split this in two pieces the before and after. 4048 */ 4049 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4050 if (nrsm == NULL) { 4051 /* 4052 * failed XXXrrs what can we do but loose the sack 4053 * info? 4054 */ 4055 goto out; 4056 } 4057 nrsm->r_start = start; 4058 nrsm->r_rtr_bytes = 0; 4059 nrsm->r_end = rsm->r_end; 4060 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4061 nrsm->r_flags = rsm->r_flags; 4062 nrsm->r_sndcnt = rsm->r_sndcnt; 4063 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4064 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4065 } 4066 rsm->r_end = nrsm->r_start; 4067 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4068 if (rsm->r_in_tmap) { 4069 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4070 nrsm->r_in_tmap = 1; 4071 } 4072 rsm->r_flags &= (~RACK_HAS_FIN); 4073 rsm = nrsm; 4074 } 4075 if (SEQ_GEQ(end, rsm->r_end)) { 4076 /* 4077 * The end of this block is either beyond this guy or right 4078 * at this guy. 4079 */ 4080 4081 if ((rsm->r_flags & RACK_ACKED) == 0) { 4082 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4083 changed += (rsm->r_end - rsm->r_start); 4084 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4085 rack_log_sack_passed(tp, rack, rsm); 4086 /* Is Reordering occuring? */ 4087 if (rsm->r_flags & RACK_SACK_PASSED) { 4088 counter_u64_add(rack_reorder_seen, 1); 4089 rack->r_ctl.rc_reorder_ts = cts; 4090 } 4091 rsm->r_flags |= RACK_ACKED; 4092 rsm->r_flags &= ~RACK_TLP; 4093 if (rsm->r_in_tmap) { 4094 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4095 rsm->r_in_tmap = 0; 4096 } 4097 } 4098 if (end == rsm->r_end) { 4099 /* This block only - done */ 4100 goto out; 4101 } 4102 /* There is more not coverend by this rsm move on */ 4103 start = rsm->r_end; 4104 nrsm = TAILQ_NEXT(rsm, r_next); 4105 rsm = nrsm; 4106 times = 0; 4107 goto do_rest_ofb; 4108 } 4109 /* Ok we need to split off this one at the tail */ 4110 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); 4111 if (nrsm == NULL) { 4112 /* failed rrs what can we do but loose the sack info? */ 4113 goto out; 4114 } 4115 /* Clone it */ 4116 nrsm->r_start = end; 4117 nrsm->r_end = rsm->r_end; 4118 nrsm->r_rtr_bytes = 0; 4119 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4120 nrsm->r_flags = rsm->r_flags; 4121 nrsm->r_sndcnt = rsm->r_sndcnt; 4122 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4123 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4124 } 4125 /* The sack block does not cover this guy fully */ 4126 rsm->r_flags &= (~RACK_HAS_FIN); 4127 rsm->r_end = end; 4128 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); 4129 if (rsm->r_in_tmap) { 4130 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4131 nrsm->r_in_tmap = 1; 4132 } 4133 if (rsm->r_flags & RACK_ACKED) { 4134 /* Been here done that */ 4135 goto out; 4136 } 4137 rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 4138 changed += (rsm->r_end - rsm->r_start); 4139 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 4140 rack_log_sack_passed(tp, rack, rsm); 4141 /* Is Reordering occuring? */ 4142 if (rsm->r_flags & RACK_SACK_PASSED) { 4143 counter_u64_add(rack_reorder_seen, 1); 4144 rack->r_ctl.rc_reorder_ts = cts; 4145 } 4146 rsm->r_flags |= RACK_ACKED; 4147 rsm->r_flags &= ~RACK_TLP; 4148 if (rsm->r_in_tmap) { 4149 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4150 rsm->r_in_tmap = 0; 4151 } 4152 out: 4153 if (rsm && (rsm->r_flags & RACK_ACKED)) { 4154 /* 4155 * Now can we merge this newly acked 4156 * block with either the previous or 4157 * next block? 4158 */ 4159 nrsm = TAILQ_NEXT(rsm, r_next); 4160 if (nrsm && 4161 (nrsm->r_flags & RACK_ACKED)) { 4162 /* yep this and next can be merged */ 4163 rsm = rack_merge_rsm(rack, rsm, nrsm); 4164 } 4165 /* Now what about the previous? */ 4166 nrsm = TAILQ_PREV(rsm, rack_head, r_next); 4167 if (nrsm && 4168 (nrsm->r_flags & RACK_ACKED)) { 4169 /* yep the previous and this can be merged */ 4170 rsm = rack_merge_rsm(rack, nrsm, rsm); 4171 } 4172 } 4173 if (used_ref == 0) { 4174 counter_u64_add(rack_sack_proc_all, 1); 4175 } else { 4176 counter_u64_add(rack_sack_proc_short, 1); 4177 } 4178 /* Save off where we last were */ 4179 if (rsm) 4180 rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 4181 else 4182 rack->r_ctl.rc_sacklast = NULL; 4183 *prsm = rsm; 4184 return (changed); 4185 } 4186 4187 static void inline 4188 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) 4189 { 4190 struct rack_sendmap *tmap; 4191 4192 tmap = NULL; 4193 while (rsm && (rsm->r_flags & RACK_ACKED)) { 4194 /* Its no longer sacked, mark it so */ 4195 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4196 #ifdef INVARIANTS 4197 if (rsm->r_in_tmap) { 4198 panic("rack:%p rsm:%p flags:0x%x in tmap?", 4199 rack, rsm, rsm->r_flags); 4200 } 4201 #endif 4202 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); 4203 /* Rebuild it into our tmap */ 4204 if (tmap == NULL) { 4205 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4206 tmap = rsm; 4207 } else { 4208 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); 4209 tmap = rsm; 4210 } 4211 tmap->r_in_tmap = 1; 4212 rsm = TAILQ_NEXT(rsm, r_next); 4213 } 4214 /* 4215 * Now lets possibly clear the sack filter so we start 4216 * recognizing sacks that cover this area. 4217 */ 4218 if (rack_use_sack_filter) 4219 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); 4220 4221 } 4222 4223 static void 4224 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) 4225 { 4226 uint32_t changed, last_seq, entered_recovery = 0; 4227 struct tcp_rack *rack; 4228 struct rack_sendmap *rsm; 4229 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 4230 register uint32_t th_ack; 4231 int32_t i, j, k, num_sack_blks = 0; 4232 uint32_t cts, acked, ack_point, sack_changed = 0; 4233 4234 INP_WLOCK_ASSERT(tp->t_inpcb); 4235 if (th->th_flags & TH_RST) { 4236 /* We don't log resets */ 4237 return; 4238 } 4239 rack = (struct tcp_rack *)tp->t_fb_ptr; 4240 cts = tcp_ts_getticks(); 4241 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4242 changed = 0; 4243 th_ack = th->th_ack; 4244 4245 if (SEQ_GT(th_ack, tp->snd_una)) { 4246 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); 4247 tp->t_acktime = ticks; 4248 } 4249 if (rsm && SEQ_GT(th_ack, rsm->r_start)) 4250 changed = th_ack - rsm->r_start; 4251 if (changed) { 4252 /* 4253 * The ACK point is advancing to th_ack, we must drop off 4254 * the packets in the rack log and calculate any eligble 4255 * RTT's. 4256 */ 4257 rack->r_wanted_output++; 4258 more: 4259 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4260 if (rsm == NULL) { 4261 if ((th_ack - 1) == tp->iss) { 4262 /* 4263 * For the SYN incoming case we will not 4264 * have called tcp_output for the sending of 4265 * the SYN, so there will be no map. All 4266 * other cases should probably be a panic. 4267 */ 4268 goto proc_sack; 4269 } 4270 if (tp->t_flags & TF_SENTFIN) { 4271 /* if we send a FIN we will not hav a map */ 4272 goto proc_sack; 4273 } 4274 #ifdef INVARIANTS 4275 panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n", 4276 tp, 4277 th, tp->t_state, rack, 4278 tp->snd_una, tp->snd_max, tp->snd_nxt, changed); 4279 #endif 4280 goto proc_sack; 4281 } 4282 if (SEQ_LT(th_ack, rsm->r_start)) { 4283 /* Huh map is missing this */ 4284 #ifdef INVARIANTS 4285 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", 4286 rsm->r_start, 4287 th_ack, tp->t_state, rack->r_state); 4288 #endif 4289 goto proc_sack; 4290 } 4291 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); 4292 /* Now do we consume the whole thing? */ 4293 if (SEQ_GEQ(th_ack, rsm->r_end)) { 4294 /* Its all consumed. */ 4295 uint32_t left; 4296 4297 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4298 rsm->r_rtr_bytes = 0; 4299 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 4300 if (rsm->r_in_tmap) { 4301 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); 4302 rsm->r_in_tmap = 0; 4303 } 4304 if (rack->r_ctl.rc_next == rsm) { 4305 /* scoot along the marker */ 4306 rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); 4307 } 4308 if (rsm->r_flags & RACK_ACKED) { 4309 /* 4310 * It was acked on the scoreboard -- remove 4311 * it from total 4312 */ 4313 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 4314 } else if (rsm->r_flags & RACK_SACK_PASSED) { 4315 /* 4316 * There are acked segments ACKED on the 4317 * scoreboard further up. We are seeing 4318 * reordering. 4319 */ 4320 counter_u64_add(rack_reorder_seen, 1); 4321 rsm->r_flags |= RACK_ACKED; 4322 rack->r_ctl.rc_reorder_ts = cts; 4323 } 4324 left = th_ack - rsm->r_end; 4325 if (rsm->r_rtr_cnt > 1) { 4326 /* 4327 * Technically we should make r_rtr_cnt be 4328 * monotonicly increasing and just mod it to 4329 * the timestamp it is replacing.. that way 4330 * we would have the last 3 retransmits. Now 4331 * rc_loss_count will be wrong if we 4332 * retransmit something more than 2 times in 4333 * recovery :( 4334 */ 4335 rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); 4336 } 4337 /* Free back to zone */ 4338 rack_free(rack, rsm); 4339 if (left) { 4340 goto more; 4341 } 4342 goto proc_sack; 4343 } 4344 if (rsm->r_flags & RACK_ACKED) { 4345 /* 4346 * It was acked on the scoreboard -- remove it from 4347 * total for the part being cum-acked. 4348 */ 4349 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 4350 } 4351 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 4352 rsm->r_rtr_bytes = 0; 4353 rsm->r_start = th_ack; 4354 } 4355 proc_sack: 4356 /* Check for reneging */ 4357 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 4358 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { 4359 /* 4360 * The peer has moved snd_una up to 4361 * the edge of this send, i.e. one 4362 * that it had previously acked. The only 4363 * way that can be true if the peer threw 4364 * away data (space issues) that it had 4365 * previously sacked (else it would have 4366 * given us snd_una up to (rsm->r_end). 4367 * We need to undo the acked markings here. 4368 * 4369 * Note we have to look to make sure th_ack is 4370 * our rsm->r_start in case we get an old ack 4371 * where th_ack is behind snd_una. 4372 */ 4373 rack_peer_reneges(rack, rsm, th->th_ack); 4374 } 4375 if ((to->to_flags & TOF_SACK) == 0) { 4376 /* We are done nothing left to log */ 4377 goto out; 4378 } 4379 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); 4380 if (rsm) { 4381 last_seq = rsm->r_end; 4382 } else { 4383 last_seq = tp->snd_max; 4384 } 4385 /* Sack block processing */ 4386 if (SEQ_GT(th_ack, tp->snd_una)) 4387 ack_point = th_ack; 4388 else 4389 ack_point = tp->snd_una; 4390 for (i = 0; i < to->to_nsacks; i++) { 4391 bcopy((to->to_sacks + i * TCPOLEN_SACK), 4392 &sack, sizeof(sack)); 4393 sack.start = ntohl(sack.start); 4394 sack.end = ntohl(sack.end); 4395 if (SEQ_GT(sack.end, sack.start) && 4396 SEQ_GT(sack.start, ack_point) && 4397 SEQ_LT(sack.start, tp->snd_max) && 4398 SEQ_GT(sack.end, ack_point) && 4399 SEQ_LEQ(sack.end, tp->snd_max)) { 4400 if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && 4401 (SEQ_LT(sack.end, last_seq)) && 4402 ((sack.end - sack.start) < (tp->t_maxseg / 8))) { 4403 /* 4404 * Not the last piece and its smaller than 4405 * 1/8th of a MSS. We ignore this. 4406 */ 4407 counter_u64_add(rack_runt_sacks, 1); 4408 continue; 4409 } 4410 sack_blocks[num_sack_blks] = sack; 4411 num_sack_blks++; 4412 } else if (SEQ_LEQ(sack.start, th_ack) && 4413 SEQ_LEQ(sack.end, th_ack)) { 4414 /* 4415 * Its a D-SACK block. 4416 */ 4417 /* tcp_record_dsack(sack.start, sack.end); */ 4418 } 4419 } 4420 if (num_sack_blks == 0) 4421 goto out; 4422 /* 4423 * Sort the SACK blocks so we can update the rack scoreboard with 4424 * just one pass. 4425 */ 4426 if (rack_use_sack_filter) { 4427 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, 4428 num_sack_blks, th->th_ack); 4429 ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks); 4430 } 4431 if (num_sack_blks < 2) { 4432 goto do_sack_work; 4433 } 4434 /* Sort the sacks */ 4435 for (i = 0; i < num_sack_blks; i++) { 4436 for (j = i + 1; j < num_sack_blks; j++) { 4437 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 4438 sack = sack_blocks[i]; 4439 sack_blocks[i] = sack_blocks[j]; 4440 sack_blocks[j] = sack; 4441 } 4442 } 4443 } 4444 /* 4445 * Now are any of the sack block ends the same (yes some 4446 * implememtations send these)? 4447 */ 4448 again: 4449 if (num_sack_blks > 1) { 4450 for (i = 0; i < num_sack_blks; i++) { 4451 for (j = i + 1; j < num_sack_blks; j++) { 4452 if (sack_blocks[i].end == sack_blocks[j].end) { 4453 /* 4454 * Ok these two have the same end we 4455 * want the smallest end and then 4456 * throw away the larger and start 4457 * again. 4458 */ 4459 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 4460 /* 4461 * The second block covers 4462 * more area use that 4463 */ 4464 sack_blocks[i].start = sack_blocks[j].start; 4465 } 4466 /* 4467 * Now collapse out the dup-sack and 4468 * lower the count 4469 */ 4470 for (k = (j + 1); k < num_sack_blks; k++) { 4471 sack_blocks[j].start = sack_blocks[k].start; 4472 sack_blocks[j].end = sack_blocks[k].end; 4473 j++; 4474 } 4475 num_sack_blks--; 4476 goto again; 4477 } 4478 } 4479 } 4480 } 4481 do_sack_work: 4482 rsm = rack->r_ctl.rc_sacklast; 4483 for (i = 0; i < num_sack_blks; i++) { 4484 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); 4485 if (acked) { 4486 rack->r_wanted_output++; 4487 changed += acked; 4488 sack_changed += acked; 4489 } 4490 } 4491 out: 4492 if (changed) { 4493 /* Something changed cancel the rack timer */ 4494 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4495 } 4496 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 4497 /* 4498 * Ok we have a high probability that we need to go in to 4499 * recovery since we have data sack'd 4500 */ 4501 struct rack_sendmap *rsm; 4502 uint32_t tsused; 4503 4504 tsused = tcp_ts_getticks(); 4505 rsm = tcp_rack_output(tp, rack, tsused); 4506 if (rsm) { 4507 /* Enter recovery */ 4508 rack->r_ctl.rc_rsm_start = rsm->r_start; 4509 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 4510 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 4511 entered_recovery = 1; 4512 rack_cong_signal(tp, NULL, CC_NDUPACK); 4513 /* 4514 * When we enter recovery we need to assure we send 4515 * one packet. 4516 */ 4517 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 4518 rack->r_timer_override = 1; 4519 } 4520 } 4521 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 4522 /* Deal with changed an PRR here (in recovery only) */ 4523 uint32_t pipe, snd_una; 4524 4525 rack->r_ctl.rc_prr_delivered += changed; 4526 /* Compute prr_sndcnt */ 4527 if (SEQ_GT(tp->snd_una, th_ack)) { 4528 snd_una = tp->snd_una; 4529 } else { 4530 snd_una = th_ack; 4531 } 4532 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 4533 if (pipe > tp->snd_ssthresh) { 4534 long sndcnt; 4535 4536 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; 4537 if (rack->r_ctl.rc_prr_recovery_fs > 0) 4538 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; 4539 else { 4540 rack->r_ctl.rc_prr_sndcnt = 0; 4541 sndcnt = 0; 4542 } 4543 sndcnt++; 4544 if (sndcnt > (long)rack->r_ctl.rc_prr_out) 4545 sndcnt -= rack->r_ctl.rc_prr_out; 4546 else 4547 sndcnt = 0; 4548 rack->r_ctl.rc_prr_sndcnt = sndcnt; 4549 } else { 4550 uint32_t limit; 4551 4552 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) 4553 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); 4554 else 4555 limit = 0; 4556 if (changed > limit) 4557 limit = changed; 4558 limit += tp->t_maxseg; 4559 if (tp->snd_ssthresh > pipe) { 4560 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); 4561 } else { 4562 rack->r_ctl.rc_prr_sndcnt = min(0, limit); 4563 } 4564 } 4565 if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { 4566 rack->r_timer_override = 1; 4567 } 4568 } 4569 } 4570 4571 /* 4572 * Return value of 1, we do not need to call rack_process_data(). 4573 * return value of 0, rack_process_data can be called. 4574 * For ret_val if its 0 the TCP is locked, if its non-zero 4575 * its unlocked and probably unsafe to touch the TCB. 4576 */ 4577 static int 4578 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 4579 struct tcpcb *tp, struct tcpopt *to, 4580 uint32_t tiwin, int32_t tlen, 4581 int32_t * ofia, int32_t thflags, int32_t * ret_val) 4582 { 4583 int32_t ourfinisacked = 0; 4584 int32_t nsegs, acked_amount; 4585 int32_t acked; 4586 struct mbuf *mfree; 4587 struct tcp_rack *rack; 4588 int32_t recovery = 0; 4589 4590 rack = (struct tcp_rack *)tp->t_fb_ptr; 4591 if (SEQ_GT(th->th_ack, tp->snd_max)) { 4592 rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 4593 return (1); 4594 } 4595 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 4596 rack_log_ack(tp, to, th); 4597 } 4598 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 4599 /* 4600 * Old ack, behind (or duplicate to) the last one rcv'd 4601 * Note: Should mark reordering is occuring! We should also 4602 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, 4603 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no 4604 * retran and> ack 3 4605 */ 4606 return (0); 4607 } 4608 /* 4609 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 4610 * something we sent. 4611 */ 4612 if (tp->t_flags & TF_NEEDSYN) { 4613 /* 4614 * T/TCP: Connection was half-synchronized, and our SYN has 4615 * been ACK'd (so connection is now fully synchronized). Go 4616 * to non-starred state, increment snd_una for ACK of SYN, 4617 * and check if we can do window scaling. 4618 */ 4619 tp->t_flags &= ~TF_NEEDSYN; 4620 tp->snd_una++; 4621 /* Do window scaling? */ 4622 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 4623 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 4624 tp->rcv_scale = tp->request_r_scale; 4625 /* Send window already scaled. */ 4626 } 4627 } 4628 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4629 INP_WLOCK_ASSERT(tp->t_inpcb); 4630 4631 acked = BYTES_THIS_ACK(tp, th); 4632 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 4633 TCPSTAT_ADD(tcps_rcvackbyte, acked); 4634 4635 /* 4636 * If we just performed our first retransmit, and the ACK arrives 4637 * within our recovery window, then it was a mistake to do the 4638 * retransmit in the first place. Recover our original cwnd and 4639 * ssthresh, and proceed to transmit where we left off. 4640 */ 4641 if (tp->t_flags & TF_PREVVALID) { 4642 tp->t_flags &= ~TF_PREVVALID; 4643 if (tp->t_rxtshift == 1 && 4644 (int)(ticks - tp->t_badrxtwin) < 0) 4645 rack_cong_signal(tp, th, CC_RTO_ERR); 4646 } 4647 /* 4648 * If we have a timestamp reply, update smoothed round trip time. If 4649 * no timestamp is present but transmit timer is running and timed 4650 * sequence number was acked, update smoothed round trip time. Since 4651 * we now have an rtt measurement, cancel the timer backoff (cf., 4652 * Phil Karn's retransmit alg.). Recompute the initial retransmit 4653 * timer. 4654 * 4655 * Some boxes send broken timestamp replies during the SYN+ACK 4656 * phase, ignore timestamps of 0 or we could calculate a huge RTT 4657 * and blow up the retransmit timer. 4658 */ 4659 /* 4660 * If all outstanding data is acked, stop retransmit timer and 4661 * remember to restart (more output or persist). If there is more 4662 * data to be acked, restart retransmit timer, using current 4663 * (possibly backed-off) value. 4664 */ 4665 if (th->th_ack == tp->snd_max) { 4666 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4667 rack->r_wanted_output++; 4668 } 4669 /* 4670 * If no data (only SYN) was ACK'd, skip rest of ACK processing. 4671 */ 4672 if (acked == 0) { 4673 if (ofia) 4674 *ofia = ourfinisacked; 4675 return (0); 4676 } 4677 if (rack->r_ctl.rc_early_recovery) { 4678 if (IN_RECOVERY(tp->t_flags)) { 4679 if (SEQ_LT(th->th_ack, tp->snd_recover) && 4680 (SEQ_LT(th->th_ack, tp->snd_max))) { 4681 tcp_rack_partialack(tp, th); 4682 } else { 4683 rack_post_recovery(tp, th); 4684 recovery = 1; 4685 } 4686 } 4687 } 4688 /* 4689 * Let the congestion control algorithm update congestion control 4690 * related information. This typically means increasing the 4691 * congestion window. 4692 */ 4693 rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); 4694 SOCKBUF_LOCK(&so->so_snd); 4695 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 4696 tp->snd_wnd -= acked_amount; 4697 mfree = sbcut_locked(&so->so_snd, acked_amount); 4698 if ((sbused(&so->so_snd) == 0) && 4699 (acked > acked_amount) && 4700 (tp->t_state >= TCPS_FIN_WAIT_1)) { 4701 ourfinisacked = 1; 4702 } 4703 /* NB: sowwakeup_locked() does an implicit unlock. */ 4704 sowwakeup_locked(so); 4705 m_freem(mfree); 4706 if (rack->r_ctl.rc_early_recovery == 0) { 4707 if (IN_RECOVERY(tp->t_flags)) { 4708 if (SEQ_LT(th->th_ack, tp->snd_recover) && 4709 (SEQ_LT(th->th_ack, tp->snd_max))) { 4710 tcp_rack_partialack(tp, th); 4711 } else { 4712 rack_post_recovery(tp, th); 4713 } 4714 } 4715 } 4716 tp->snd_una = th->th_ack; 4717 if (SEQ_GT(tp->snd_una, tp->snd_recover)) 4718 tp->snd_recover = tp->snd_una; 4719 4720 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { 4721 tp->snd_nxt = tp->snd_una; 4722 } 4723 if (tp->snd_una == tp->snd_max) { 4724 /* Nothing left outstanding */ 4725 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 4726 tp->t_acktime = 0; 4727 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4728 /* Set need output so persist might get set */ 4729 rack->r_wanted_output++; 4730 if (rack_use_sack_filter) 4731 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 4732 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 4733 (sbavail(&so->so_snd) == 0) && 4734 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 4735 /* 4736 * The socket was gone and the 4737 * peer sent data, time to 4738 * reset him. 4739 */ 4740 *ret_val = 1; 4741 tp = tcp_close(tp); 4742 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 4743 return (1); 4744 } 4745 } 4746 if (ofia) 4747 *ofia = ourfinisacked; 4748 return (0); 4749 } 4750 4751 4752 /* 4753 * Return value of 1, the TCB is unlocked and most 4754 * likely gone, return value of 0, the TCP is still 4755 * locked. 4756 */ 4757 static int 4758 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 4759 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 4760 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 4761 { 4762 /* 4763 * Update window information. Don't look at window if no ACK: TAC's 4764 * send garbage on first SYN. 4765 */ 4766 int32_t nsegs; 4767 #ifdef TCP_RFC7413 4768 int32_t tfo_syn; 4769 #else 4770 #define tfo_syn (FALSE) 4771 #endif 4772 struct tcp_rack *rack; 4773 4774 rack = (struct tcp_rack *)tp->t_fb_ptr; 4775 INP_WLOCK_ASSERT(tp->t_inpcb); 4776 nsegs = max(1, m->m_pkthdr.lro_nsegs); 4777 if ((thflags & TH_ACK) && 4778 (SEQ_LT(tp->snd_wl1, th->th_seq) || 4779 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 4780 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 4781 /* keep track of pure window updates */ 4782 if (tlen == 0 && 4783 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 4784 TCPSTAT_INC(tcps_rcvwinupd); 4785 tp->snd_wnd = tiwin; 4786 tp->snd_wl1 = th->th_seq; 4787 tp->snd_wl2 = th->th_ack; 4788 if (tp->snd_wnd > tp->max_sndwnd) 4789 tp->max_sndwnd = tp->snd_wnd; 4790 rack->r_wanted_output++; 4791 } else if (thflags & TH_ACK) { 4792 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 4793 tp->snd_wnd = tiwin; 4794 tp->snd_wl1 = th->th_seq; 4795 tp->snd_wl2 = th->th_ack; 4796 } 4797 } 4798 /* Was persist timer active and now we have window space? */ 4799 if ((rack->rc_in_persist != 0) && tp->snd_wnd) { 4800 rack_exit_persist(tp, rack); 4801 tp->snd_nxt = tp->snd_max; 4802 /* Make sure we output to start the timer */ 4803 rack->r_wanted_output++; 4804 } 4805 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 4806 m_freem(m); 4807 return (0); 4808 } 4809 /* 4810 * Process segments with URG. 4811 */ 4812 if ((thflags & TH_URG) && th->th_urp && 4813 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4814 /* 4815 * This is a kludge, but if we receive and accept random 4816 * urgent pointers, we'll crash in soreceive. It's hard to 4817 * imagine someone actually wanting to send this much urgent 4818 * data. 4819 */ 4820 SOCKBUF_LOCK(&so->so_rcv); 4821 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { 4822 th->th_urp = 0; /* XXX */ 4823 thflags &= ~TH_URG; /* XXX */ 4824 SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ 4825 goto dodata; /* XXX */ 4826 } 4827 /* 4828 * If this segment advances the known urgent pointer, then 4829 * mark the data stream. This should not happen in 4830 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a 4831 * FIN has been received from the remote side. In these 4832 * states we ignore the URG. 4833 * 4834 * According to RFC961 (Assigned Protocols), the urgent 4835 * pointer points to the last octet of urgent data. We 4836 * continue, however, to consider it to indicate the first 4837 * octet of data past the urgent section as the original 4838 * spec states (in one of two places). 4839 */ 4840 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { 4841 tp->rcv_up = th->th_seq + th->th_urp; 4842 so->so_oobmark = sbavail(&so->so_rcv) + 4843 (tp->rcv_up - tp->rcv_nxt) - 1; 4844 if (so->so_oobmark == 0) 4845 so->so_rcv.sb_state |= SBS_RCVATMARK; 4846 sohasoutofband(so); 4847 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); 4848 } 4849 SOCKBUF_UNLOCK(&so->so_rcv); 4850 /* 4851 * Remove out of band data so doesn't get presented to user. 4852 * This can happen independent of advancing the URG pointer, 4853 * but if two URG's are pending at once, some out-of-band 4854 * data may creep in... ick. 4855 */ 4856 if (th->th_urp <= (uint32_t) tlen && 4857 !(so->so_options & SO_OOBINLINE)) { 4858 /* hdr drop is delayed */ 4859 tcp_pulloutofband(so, th, m, drop_hdrlen); 4860 } 4861 } else { 4862 /* 4863 * If no out of band data is expected, pull receive urgent 4864 * pointer along with the receive window. 4865 */ 4866 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) 4867 tp->rcv_up = tp->rcv_nxt; 4868 } 4869 dodata: /* XXX */ 4870 INP_WLOCK_ASSERT(tp->t_inpcb); 4871 4872 /* 4873 * Process the segment text, merging it into the TCP sequencing 4874 * queue, and arranging for acknowledgment of receipt if necessary. 4875 * This process logically involves adjusting tp->rcv_wnd as data is 4876 * presented to the user (this happens in tcp_usrreq.c, case 4877 * PRU_RCVD). If a FIN has already been received on this connection 4878 * then we just ignore the text. 4879 */ 4880 #ifdef TCP_RFC7413 4881 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 4882 (tp->t_flags & TF_FASTOPEN)); 4883 #endif 4884 if ((tlen || (thflags & TH_FIN) || tfo_syn) && 4885 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4886 tcp_seq save_start = th->th_seq; 4887 tcp_seq save_rnxt = tp->rcv_nxt; 4888 int save_tlen = tlen; 4889 4890 m_adj(m, drop_hdrlen); /* delayed header drop */ 4891 /* 4892 * Insert segment which includes th into TCP reassembly 4893 * queue with control block tp. Set thflags to whether 4894 * reassembly now includes a segment with FIN. This handles 4895 * the common case inline (segment is the next to be 4896 * received on an established connection, and the queue is 4897 * empty), avoiding linkage into and removal from the queue 4898 * and repetition of various conversions. Set DELACK for 4899 * segments received in order, but ack immediately when 4900 * segments are out of order (so fast retransmit can work). 4901 */ 4902 if (th->th_seq == tp->rcv_nxt && 4903 SEGQ_EMPTY(tp) && 4904 (TCPS_HAVEESTABLISHED(tp->t_state) || 4905 tfo_syn)) { 4906 if (DELAY_ACK(tp, tlen) || tfo_syn) { 4907 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4908 tp->t_flags |= TF_DELACK; 4909 } else { 4910 rack->r_wanted_output++; 4911 tp->t_flags |= TF_ACKNOW; 4912 } 4913 tp->rcv_nxt += tlen; 4914 thflags = th->th_flags & TH_FIN; 4915 TCPSTAT_ADD(tcps_rcvpack, nsegs); 4916 TCPSTAT_ADD(tcps_rcvbyte, tlen); 4917 SOCKBUF_LOCK(&so->so_rcv); 4918 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 4919 m_freem(m); 4920 else 4921 sbappendstream_locked(&so->so_rcv, m, 0); 4922 /* NB: sorwakeup_locked() does an implicit unlock. */ 4923 sorwakeup_locked(so); 4924 } else { 4925 /* 4926 * XXX: Due to the header drop above "th" is 4927 * theoretically invalid by now. Fortunately 4928 * m_adj() doesn't actually frees any mbufs when 4929 * trimming from the head. 4930 */ 4931 tcp_seq temp = save_start; 4932 thflags = tcp_reass(tp, th, &temp, &tlen, m); 4933 tp->t_flags |= TF_ACKNOW; 4934 } 4935 if (((tlen == 0) && (save_tlen > 0) && 4936 (SEQ_LT(save_start, save_rnxt)))) { 4937 /* 4938 * DSACK actually handled in the fastpath 4939 * above. 4940 */ 4941 tcp_update_sack_list(tp, save_start, save_start + save_tlen); 4942 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 4943 /* 4944 * Cleaning sackblks by using zero length 4945 * update. 4946 */ 4947 tcp_update_sack_list(tp, save_start, save_start); 4948 } else if ((tlen > 0) && (tlen >= save_tlen)) { 4949 /* Update of sackblks. */ 4950 tcp_update_sack_list(tp, save_start, save_start + save_tlen); 4951 } else if (tlen > 0) { 4952 tcp_update_sack_list(tp, save_start, save_start+tlen); 4953 } 4954 } else { 4955 m_freem(m); 4956 thflags &= ~TH_FIN; 4957 } 4958 4959 /* 4960 * If FIN is received ACK the FIN and let the user know that the 4961 * connection is closing. 4962 */ 4963 if (thflags & TH_FIN) { 4964 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 4965 socantrcvmore(so); 4966 /* 4967 * If connection is half-synchronized (ie NEEDSYN 4968 * flag on) then delay ACK, so it may be piggybacked 4969 * when SYN is sent. Otherwise, since we received a 4970 * FIN then no more input can be expected, send ACK 4971 * now. 4972 */ 4973 if (tp->t_flags & TF_NEEDSYN) { 4974 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4975 tp->t_flags |= TF_DELACK; 4976 } else { 4977 tp->t_flags |= TF_ACKNOW; 4978 } 4979 tp->rcv_nxt++; 4980 } 4981 switch (tp->t_state) { 4982 4983 /* 4984 * In SYN_RECEIVED and ESTABLISHED STATES enter the 4985 * CLOSE_WAIT state. 4986 */ 4987 case TCPS_SYN_RECEIVED: 4988 tp->t_starttime = ticks; 4989 /* FALLTHROUGH */ 4990 case TCPS_ESTABLISHED: 4991 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 4992 tcp_state_change(tp, TCPS_CLOSE_WAIT); 4993 break; 4994 4995 /* 4996 * If still in FIN_WAIT_1 STATE FIN has not been 4997 * acked so enter the CLOSING state. 4998 */ 4999 case TCPS_FIN_WAIT_1: 5000 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5001 tcp_state_change(tp, TCPS_CLOSING); 5002 break; 5003 5004 /* 5005 * In FIN_WAIT_2 state enter the TIME_WAIT state, 5006 * starting the time-wait timer, turning off the 5007 * other standard timers. 5008 */ 5009 case TCPS_FIN_WAIT_2: 5010 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5011 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5012 tcp_twstart(tp); 5013 return (1); 5014 } 5015 } 5016 /* 5017 * Return any desired output. 5018 */ 5019 if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { 5020 rack->r_wanted_output++; 5021 } 5022 INP_WLOCK_ASSERT(tp->t_inpcb); 5023 return (0); 5024 } 5025 5026 /* 5027 * Here nothing is really faster, its just that we 5028 * have broken out the fast-data path also just like 5029 * the fast-ack. 5030 */ 5031 static int 5032 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 5033 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5034 uint32_t tiwin, int32_t nxt_pkt) 5035 { 5036 int32_t nsegs; 5037 int32_t newsize = 0; /* automatic sockbuf scaling */ 5038 struct tcp_rack *rack; 5039 #ifdef TCPDEBUG 5040 /* 5041 * The size of tcp_saveipgen must be the size of the max ip header, 5042 * now IPv6. 5043 */ 5044 u_char tcp_saveipgen[IP6_HDR_LEN]; 5045 struct tcphdr tcp_savetcp; 5046 short ostate = 0; 5047 5048 #endif 5049 /* 5050 * If last ACK falls within this segment's sequence numbers, record 5051 * the timestamp. NOTE that the test is modified according to the 5052 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5053 */ 5054 if (__predict_false(th->th_seq != tp->rcv_nxt)) { 5055 return (0); 5056 } 5057 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5058 return (0); 5059 } 5060 if (tiwin && tiwin != tp->snd_wnd) { 5061 return (0); 5062 } 5063 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 5064 return (0); 5065 } 5066 if (__predict_false((to->to_flags & TOF_TS) && 5067 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 5068 return (0); 5069 } 5070 if (__predict_false((th->th_ack != tp->snd_una))) { 5071 return (0); 5072 } 5073 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 5074 return (0); 5075 } 5076 if ((to->to_flags & TOF_TS) != 0 && 5077 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5078 tp->ts_recent_age = tcp_ts_getticks(); 5079 tp->ts_recent = to->to_tsval; 5080 } 5081 rack = (struct tcp_rack *)tp->t_fb_ptr; 5082 /* 5083 * This is a pure, in-sequence data packet with nothing on the 5084 * reassembly queue and we have enough buffer space to take it. 5085 */ 5086 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5087 5088 5089 /* Clean receiver SACK report if present */ 5090 if (tp->rcv_numsacks) 5091 tcp_clean_sackreport(tp); 5092 TCPSTAT_INC(tcps_preddat); 5093 tp->rcv_nxt += tlen; 5094 /* 5095 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 5096 */ 5097 tp->snd_wl1 = th->th_seq; 5098 /* 5099 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 5100 */ 5101 tp->rcv_up = tp->rcv_nxt; 5102 TCPSTAT_ADD(tcps_rcvpack, nsegs); 5103 TCPSTAT_ADD(tcps_rcvbyte, tlen); 5104 #ifdef TCPDEBUG 5105 if (so->so_options & SO_DEBUG) 5106 tcp_trace(TA_INPUT, ostate, tp, 5107 (void *)tcp_saveipgen, &tcp_savetcp, 0); 5108 #endif 5109 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 5110 5111 /* Add data to socket buffer. */ 5112 SOCKBUF_LOCK(&so->so_rcv); 5113 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5114 m_freem(m); 5115 } else { 5116 /* 5117 * Set new socket buffer size. Give up when limit is 5118 * reached. 5119 */ 5120 if (newsize) 5121 if (!sbreserve_locked(&so->so_rcv, 5122 newsize, so, NULL)) 5123 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 5124 m_adj(m, drop_hdrlen); /* delayed header drop */ 5125 sbappendstream_locked(&so->so_rcv, m, 0); 5126 rack_calc_rwin(so, tp); 5127 } 5128 /* NB: sorwakeup_locked() does an implicit unlock. */ 5129 sorwakeup_locked(so); 5130 if (DELAY_ACK(tp, tlen)) { 5131 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5132 tp->t_flags |= TF_DELACK; 5133 } else { 5134 tp->t_flags |= TF_ACKNOW; 5135 rack->r_wanted_output++; 5136 } 5137 if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 5138 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); 5139 return (1); 5140 } 5141 5142 /* 5143 * This subfunction is used to try to highly optimize the 5144 * fast path. We again allow window updates that are 5145 * in sequence to remain in the fast-path. We also add 5146 * in the __predict's to attempt to help the compiler. 5147 * Note that if we return a 0, then we can *not* process 5148 * it and the caller should push the packet into the 5149 * slow-path. 5150 */ 5151 static int 5152 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 5153 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5154 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) 5155 { 5156 int32_t acked; 5157 int32_t nsegs; 5158 5159 #ifdef TCPDEBUG 5160 /* 5161 * The size of tcp_saveipgen must be the size of the max ip header, 5162 * now IPv6. 5163 */ 5164 u_char tcp_saveipgen[IP6_HDR_LEN]; 5165 struct tcphdr tcp_savetcp; 5166 short ostate = 0; 5167 5168 #endif 5169 struct tcp_rack *rack; 5170 5171 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 5172 /* Old ack, behind (or duplicate to) the last one rcv'd */ 5173 return (0); 5174 } 5175 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 5176 /* Above what we have sent? */ 5177 return (0); 5178 } 5179 if (__predict_false(tp->snd_nxt != tp->snd_max)) { 5180 /* We are retransmitting */ 5181 return (0); 5182 } 5183 if (__predict_false(tiwin == 0)) { 5184 /* zero window */ 5185 return (0); 5186 } 5187 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 5188 /* We need a SYN or a FIN, unlikely.. */ 5189 return (0); 5190 } 5191 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 5192 /* Timestamp is behind .. old ack with seq wrap? */ 5193 return (0); 5194 } 5195 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 5196 /* Still recovering */ 5197 return (0); 5198 } 5199 rack = (struct tcp_rack *)tp->t_fb_ptr; 5200 if (rack->r_ctl.rc_sacked) { 5201 /* We have sack holes on our scoreboard */ 5202 return (0); 5203 } 5204 /* Ok if we reach here, we can process a fast-ack */ 5205 nsegs = max(1, m->m_pkthdr.lro_nsegs); 5206 rack_log_ack(tp, to, th); 5207 /* Did the window get updated? */ 5208 if (tiwin != tp->snd_wnd) { 5209 tp->snd_wnd = tiwin; 5210 tp->snd_wl1 = th->th_seq; 5211 if (tp->snd_wnd > tp->max_sndwnd) 5212 tp->max_sndwnd = tp->snd_wnd; 5213 } 5214 if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { 5215 rack_exit_persist(tp, rack); 5216 } 5217 /* 5218 * If last ACK falls within this segment's sequence numbers, record 5219 * the timestamp. NOTE that the test is modified according to the 5220 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 5221 */ 5222 if ((to->to_flags & TOF_TS) != 0 && 5223 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 5224 tp->ts_recent_age = tcp_ts_getticks(); 5225 tp->ts_recent = to->to_tsval; 5226 } 5227 /* 5228 * This is a pure ack for outstanding data. 5229 */ 5230 TCPSTAT_INC(tcps_predack); 5231 5232 /* 5233 * "bad retransmit" recovery. 5234 */ 5235 if (tp->t_flags & TF_PREVVALID) { 5236 tp->t_flags &= ~TF_PREVVALID; 5237 if (tp->t_rxtshift == 1 && 5238 (int)(ticks - tp->t_badrxtwin) < 0) 5239 rack_cong_signal(tp, th, CC_RTO_ERR); 5240 } 5241 /* 5242 * Recalculate the transmit timer / rtt. 5243 * 5244 * Some boxes send broken timestamp replies during the SYN+ACK 5245 * phase, ignore timestamps of 0 or we could calculate a huge RTT 5246 * and blow up the retransmit timer. 5247 */ 5248 acked = BYTES_THIS_ACK(tp, th); 5249 5250 #ifdef TCP_HHOOK 5251 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 5252 hhook_run_tcp_est_in(tp, th, to); 5253 #endif 5254 5255 TCPSTAT_ADD(tcps_rcvackpack, nsegs); 5256 TCPSTAT_ADD(tcps_rcvackbyte, acked); 5257 sbdrop(&so->so_snd, acked); 5258 /* 5259 * Let the congestion control algorithm update congestion control 5260 * related information. This typically means increasing the 5261 * congestion window. 5262 */ 5263 rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); 5264 5265 tp->snd_una = th->th_ack; 5266 /* 5267 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 5268 */ 5269 tp->snd_wl2 = th->th_ack; 5270 tp->t_dupacks = 0; 5271 m_freem(m); 5272 /* ND6_HINT(tp); *//* Some progress has been made. */ 5273 5274 /* 5275 * If all outstanding data are acked, stop retransmit timer, 5276 * otherwise restart timer using current (possibly backed-off) 5277 * value. If process is waiting for space, wakeup/selwakeup/signal. 5278 * If data are ready to send, let tcp_output decide between more 5279 * output or persist. 5280 */ 5281 #ifdef TCPDEBUG 5282 if (so->so_options & SO_DEBUG) 5283 tcp_trace(TA_INPUT, ostate, tp, 5284 (void *)tcp_saveipgen, 5285 &tcp_savetcp, 0); 5286 #endif 5287 if (tp->snd_una == tp->snd_max) { 5288 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); 5289 tp->t_acktime = 0; 5290 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 5291 } 5292 /* Wake up the socket if we have room to write more */ 5293 sowwakeup(so); 5294 if (sbavail(&so->so_snd)) { 5295 rack->r_wanted_output++; 5296 } 5297 return (1); 5298 } 5299 5300 /* 5301 * Return value of 1, the TCB is unlocked and most 5302 * likely gone, return value of 0, the TCP is still 5303 * locked. 5304 */ 5305 static int 5306 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 5307 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5308 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5309 { 5310 int32_t ret_val = 0; 5311 int32_t todrop; 5312 int32_t ourfinisacked = 0; 5313 5314 rack_calc_rwin(so, tp); 5315 /* 5316 * If the state is SYN_SENT: if seg contains an ACK, but not for our 5317 * SYN, drop the input. if seg contains a RST, then drop the 5318 * connection. if seg does not contain SYN, then drop it. Otherwise 5319 * this is an acceptable SYN segment initialize tp->rcv_nxt and 5320 * tp->irs if seg contains ack then advance tp->snd_una if seg 5321 * contains an ECE and ECN support is enabled, the stream is ECN 5322 * capable. if SYN has been acked change to ESTABLISHED else 5323 * SYN_RCVD state arrange for segment to be acked (eventually) 5324 * continue processing rest of data/controls, beginning with URG 5325 */ 5326 if ((thflags & TH_ACK) && 5327 (SEQ_LEQ(th->th_ack, tp->iss) || 5328 SEQ_GT(th->th_ack, tp->snd_max))) { 5329 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5330 return (1); 5331 } 5332 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 5333 TCP_PROBE5(connect__refused, NULL, tp, 5334 mtod(m, const char *), tp, th); 5335 tp = tcp_drop(tp, ECONNREFUSED); 5336 rack_do_drop(m, tp); 5337 return (1); 5338 } 5339 if (thflags & TH_RST) { 5340 rack_do_drop(m, tp); 5341 return (1); 5342 } 5343 if (!(thflags & TH_SYN)) { 5344 rack_do_drop(m, tp); 5345 return (1); 5346 } 5347 tp->irs = th->th_seq; 5348 tcp_rcvseqinit(tp); 5349 if (thflags & TH_ACK) { 5350 TCPSTAT_INC(tcps_connects); 5351 soisconnected(so); 5352 #ifdef MAC 5353 mac_socketpeer_set_from_mbuf(m, so); 5354 #endif 5355 /* Do window scaling on this connection? */ 5356 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5357 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5358 tp->rcv_scale = tp->request_r_scale; 5359 } 5360 tp->rcv_adv += min(tp->rcv_wnd, 5361 TCP_MAXWIN << tp->rcv_scale); 5362 /* 5363 * If there's data, delay ACK; if there's also a FIN ACKNOW 5364 * will be turned on later. 5365 */ 5366 if (DELAY_ACK(tp, tlen) && tlen != 0) { 5367 rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, 5368 ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); 5369 tp->t_flags |= TF_DELACK; 5370 } else { 5371 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; 5372 tp->t_flags |= TF_ACKNOW; 5373 } 5374 5375 if ((thflags & TH_ECE) && V_tcp_do_ecn) { 5376 tp->t_flags |= TF_ECN_PERMIT; 5377 TCPSTAT_INC(tcps_ecn_shs); 5378 } 5379 /* 5380 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 5381 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 5382 */ 5383 tp->t_starttime = ticks; 5384 if (tp->t_flags & TF_NEEDFIN) { 5385 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5386 tp->t_flags &= ~TF_NEEDFIN; 5387 thflags &= ~TH_SYN; 5388 } else { 5389 tcp_state_change(tp, TCPS_ESTABLISHED); 5390 TCP_PROBE5(connect__established, NULL, tp, 5391 mtod(m, const char *), tp, th); 5392 cc_conn_init(tp); 5393 } 5394 } else { 5395 /* 5396 * Received initial SYN in SYN-SENT[*] state => simultaneous 5397 * open. If segment contains CC option and there is a 5398 * cached CC, apply TAO test. If it succeeds, connection is * 5399 * half-synchronized. Otherwise, do 3-way handshake: 5400 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 5401 * there was no CC option, clear cached CC value. 5402 */ 5403 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 5404 tcp_state_change(tp, TCPS_SYN_RECEIVED); 5405 } 5406 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5407 INP_WLOCK_ASSERT(tp->t_inpcb); 5408 /* 5409 * Advance th->th_seq to correspond to first data byte. If data, 5410 * trim to stay within window, dropping FIN if necessary. 5411 */ 5412 th->th_seq++; 5413 if (tlen > tp->rcv_wnd) { 5414 todrop = tlen - tp->rcv_wnd; 5415 m_adj(m, -todrop); 5416 tlen = tp->rcv_wnd; 5417 thflags &= ~TH_FIN; 5418 TCPSTAT_INC(tcps_rcvpackafterwin); 5419 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 5420 } 5421 tp->snd_wl1 = th->th_seq - 1; 5422 tp->rcv_up = th->th_seq; 5423 /* 5424 * Client side of transaction: already sent SYN and data. If the 5425 * remote host used T/TCP to validate the SYN, our data will be 5426 * ACK'd; if so, enter normal data segment processing in the middle 5427 * of step 5, ack processing. Otherwise, goto step 6. 5428 */ 5429 if (thflags & TH_ACK) { 5430 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 5431 return (ret_val); 5432 /* We may have changed to FIN_WAIT_1 above */ 5433 if (tp->t_state == TCPS_FIN_WAIT_1) { 5434 /* 5435 * In FIN_WAIT_1 STATE in addition to the processing 5436 * for the ESTABLISHED state if our FIN is now 5437 * acknowledged then enter FIN_WAIT_2. 5438 */ 5439 if (ourfinisacked) { 5440 /* 5441 * If we can't receive any more data, then 5442 * closing user can proceed. Starting the 5443 * timer is contrary to the specification, 5444 * but if we don't get a FIN we'll hang 5445 * forever. 5446 * 5447 * XXXjl: we should release the tp also, and 5448 * use a compressed state. 5449 */ 5450 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5451 soisdisconnected(so); 5452 tcp_timer_activate(tp, TT_2MSL, 5453 (tcp_fast_finwait2_recycle ? 5454 tcp_finwait2_timeout : 5455 TP_MAXIDLE(tp))); 5456 } 5457 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5458 } 5459 } 5460 } 5461 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5462 tiwin, thflags, nxt_pkt)); 5463 } 5464 5465 /* 5466 * Return value of 1, the TCB is unlocked and most 5467 * likely gone, return value of 0, the TCP is still 5468 * locked. 5469 */ 5470 static int 5471 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 5472 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5473 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5474 { 5475 int32_t ret_val = 0; 5476 int32_t ourfinisacked = 0; 5477 5478 rack_calc_rwin(so, tp); 5479 5480 if ((thflags & TH_ACK) && 5481 (SEQ_LEQ(th->th_ack, tp->snd_una) || 5482 SEQ_GT(th->th_ack, tp->snd_max))) { 5483 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5484 return (1); 5485 } 5486 #ifdef TCP_RFC7413 5487 if (tp->t_flags & TF_FASTOPEN) { 5488 /* 5489 * When a TFO connection is in SYN_RECEIVED, the only valid 5490 * packets are the initial SYN, a retransmit/copy of the 5491 * initial SYN (possibly with a subset of the original 5492 * data), a valid ACK, a FIN, or a RST. 5493 */ 5494 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 5495 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5496 return (1); 5497 } else if (thflags & TH_SYN) { 5498 /* non-initial SYN is ignored */ 5499 struct tcp_rack *rack; 5500 5501 rack = (struct tcp_rack *)tp->t_fb_ptr; 5502 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 5503 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 5504 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 5505 rack_do_drop(m, NULL); 5506 return (0); 5507 } 5508 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 5509 rack_do_drop(m, NULL); 5510 return (0); 5511 } 5512 } 5513 #endif 5514 if (thflags & TH_RST) 5515 return (rack_process_rst(m, th, so, tp)); 5516 /* 5517 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5518 * synchronized state. 5519 */ 5520 if (thflags & TH_SYN) { 5521 rack_challenge_ack(m, th, tp, &ret_val); 5522 return (ret_val); 5523 } 5524 /* 5525 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5526 * it's less than ts_recent, drop it. 5527 */ 5528 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5529 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5530 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5531 return (ret_val); 5532 } 5533 /* 5534 * In the SYN-RECEIVED state, validate that the packet belongs to 5535 * this connection before trimming the data to fit the receive 5536 * window. Check the sequence number versus IRS since we know the 5537 * sequence numbers haven't wrapped. This is a partial fix for the 5538 * "LAND" DoS attack. 5539 */ 5540 if (SEQ_LT(th->th_seq, tp->irs)) { 5541 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5542 return (1); 5543 } 5544 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5545 return (ret_val); 5546 } 5547 /* 5548 * If last ACK falls within this segment's sequence numbers, record 5549 * its timestamp. NOTE: 1) That the test incorporates suggestions 5550 * from the latest proposal of the tcplw@cray.com list (Braden 5551 * 1993/04/26). 2) That updating only on newer timestamps interferes 5552 * with our earlier PAWS tests, so this check should be solely 5553 * predicated on the sequence space of this segment. 3) That we 5554 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5555 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5556 * SEG.Len, This modified check allows us to overcome RFC1323's 5557 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5558 * p.869. In such cases, we can still calculate the RTT correctly 5559 * when RCV.NXT == Last.ACK.Sent. 5560 */ 5561 if ((to->to_flags & TOF_TS) != 0 && 5562 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5563 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5564 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5565 tp->ts_recent_age = tcp_ts_getticks(); 5566 tp->ts_recent = to->to_tsval; 5567 } 5568 /* 5569 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5570 * is on (half-synchronized state), then queue data for later 5571 * processing; else drop segment and return. 5572 */ 5573 if ((thflags & TH_ACK) == 0) { 5574 #ifdef TCP_RFC7413 5575 if (tp->t_flags & TF_FASTOPEN) { 5576 tp->snd_wnd = tiwin; 5577 cc_conn_init(tp); 5578 } 5579 #endif 5580 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5581 tiwin, thflags, nxt_pkt)); 5582 } 5583 TCPSTAT_INC(tcps_connects); 5584 soisconnected(so); 5585 /* Do window scaling? */ 5586 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 5587 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 5588 tp->rcv_scale = tp->request_r_scale; 5589 tp->snd_wnd = tiwin; 5590 } 5591 /* 5592 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 5593 * FIN-WAIT-1 5594 */ 5595 tp->t_starttime = ticks; 5596 if (tp->t_flags & TF_NEEDFIN) { 5597 tcp_state_change(tp, TCPS_FIN_WAIT_1); 5598 tp->t_flags &= ~TF_NEEDFIN; 5599 } else { 5600 tcp_state_change(tp, TCPS_ESTABLISHED); 5601 TCP_PROBE5(accept__established, NULL, tp, 5602 mtod(m, const char *), tp, th); 5603 #ifdef TCP_RFC7413 5604 if (tp->t_tfo_pending) { 5605 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 5606 tp->t_tfo_pending = NULL; 5607 5608 /* 5609 * Account for the ACK of our SYN prior to regular 5610 * ACK processing below. 5611 */ 5612 tp->snd_una++; 5613 } 5614 /* 5615 * TFO connections call cc_conn_init() during SYN 5616 * processing. Calling it again here for such connections 5617 * is not harmless as it would undo the snd_cwnd reduction 5618 * that occurs when a TFO SYN|ACK is retransmitted. 5619 */ 5620 if (!(tp->t_flags & TF_FASTOPEN)) 5621 #endif 5622 cc_conn_init(tp); 5623 } 5624 /* 5625 * If segment contains data or ACK, will call tcp_reass() later; if 5626 * not, do so now to pass queued data to user. 5627 */ 5628 if (tlen == 0 && (thflags & TH_FIN) == 0) 5629 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 5630 (struct mbuf *)0); 5631 tp->snd_wl1 = th->th_seq - 1; 5632 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5633 return (ret_val); 5634 } 5635 if (tp->t_state == TCPS_FIN_WAIT_1) { 5636 /* We could have went to FIN_WAIT_1 (or EST) above */ 5637 /* 5638 * In FIN_WAIT_1 STATE in addition to the processing for the 5639 * ESTABLISHED state if our FIN is now acknowledged then 5640 * enter FIN_WAIT_2. 5641 */ 5642 if (ourfinisacked) { 5643 /* 5644 * If we can't receive any more data, then closing 5645 * user can proceed. Starting the timer is contrary 5646 * to the specification, but if we don't get a FIN 5647 * we'll hang forever. 5648 * 5649 * XXXjl: we should release the tp also, and use a 5650 * compressed state. 5651 */ 5652 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 5653 soisdisconnected(so); 5654 tcp_timer_activate(tp, TT_2MSL, 5655 (tcp_fast_finwait2_recycle ? 5656 tcp_finwait2_timeout : 5657 TP_MAXIDLE(tp))); 5658 } 5659 tcp_state_change(tp, TCPS_FIN_WAIT_2); 5660 } 5661 } 5662 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5663 tiwin, thflags, nxt_pkt)); 5664 } 5665 5666 /* 5667 * Return value of 1, the TCB is unlocked and most 5668 * likely gone, return value of 0, the TCP is still 5669 * locked. 5670 */ 5671 static int 5672 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 5673 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5674 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5675 { 5676 int32_t ret_val = 0; 5677 5678 /* 5679 * Header prediction: check for the two common cases of a 5680 * uni-directional data xfer. If the packet has no control flags, 5681 * is in-sequence, the window didn't change and we're not 5682 * retransmitting, it's a candidate. If the length is zero and the 5683 * ack moved forward, we're the sender side of the xfer. Just free 5684 * the data acked & wake any higher level process that was blocked 5685 * waiting for space. If the length is non-zero and the ack didn't 5686 * move, we're the receiver side. If we're getting packets in-order 5687 * (the reassembly queue is empty), add the data toc The socket 5688 * buffer and note that we need a delayed ack. Make sure that the 5689 * hidden state-flags are also off. Since we check for 5690 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 5691 */ 5692 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 5693 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 5694 __predict_true(SEGQ_EMPTY(tp)) && 5695 __predict_true(th->th_seq == tp->rcv_nxt)) { 5696 struct tcp_rack *rack; 5697 5698 rack = (struct tcp_rack *)tp->t_fb_ptr; 5699 if (tlen == 0) { 5700 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 5701 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { 5702 return (0); 5703 } 5704 } else { 5705 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 5706 tiwin, nxt_pkt)) { 5707 return (0); 5708 } 5709 } 5710 } 5711 rack_calc_rwin(so, tp); 5712 5713 if (thflags & TH_RST) 5714 return (rack_process_rst(m, th, so, tp)); 5715 5716 /* 5717 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5718 * synchronized state. 5719 */ 5720 if (thflags & TH_SYN) { 5721 rack_challenge_ack(m, th, tp, &ret_val); 5722 return (ret_val); 5723 } 5724 /* 5725 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5726 * it's less than ts_recent, drop it. 5727 */ 5728 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5729 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5730 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5731 return (ret_val); 5732 } 5733 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5734 return (ret_val); 5735 } 5736 /* 5737 * If last ACK falls within this segment's sequence numbers, record 5738 * its timestamp. NOTE: 1) That the test incorporates suggestions 5739 * from the latest proposal of the tcplw@cray.com list (Braden 5740 * 1993/04/26). 2) That updating only on newer timestamps interferes 5741 * with our earlier PAWS tests, so this check should be solely 5742 * predicated on the sequence space of this segment. 3) That we 5743 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5744 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5745 * SEG.Len, This modified check allows us to overcome RFC1323's 5746 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5747 * p.869. In such cases, we can still calculate the RTT correctly 5748 * when RCV.NXT == Last.ACK.Sent. 5749 */ 5750 if ((to->to_flags & TOF_TS) != 0 && 5751 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5752 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5753 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5754 tp->ts_recent_age = tcp_ts_getticks(); 5755 tp->ts_recent = to->to_tsval; 5756 } 5757 /* 5758 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5759 * is on (half-synchronized state), then queue data for later 5760 * processing; else drop segment and return. 5761 */ 5762 if ((thflags & TH_ACK) == 0) { 5763 if (tp->t_flags & TF_NEEDSYN) { 5764 5765 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5766 tiwin, thflags, nxt_pkt)); 5767 5768 } else if (tp->t_flags & TF_ACKNOW) { 5769 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5770 return (ret_val); 5771 } else { 5772 rack_do_drop(m, NULL); 5773 return (0); 5774 } 5775 } 5776 /* 5777 * Ack processing. 5778 */ 5779 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5780 return (ret_val); 5781 } 5782 if (sbavail(&so->so_snd)) { 5783 if (rack_progress_timeout_check(tp)) { 5784 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5785 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5786 return (1); 5787 } 5788 } 5789 /* State changes only happen in rack_process_data() */ 5790 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5791 tiwin, thflags, nxt_pkt)); 5792 } 5793 5794 /* 5795 * Return value of 1, the TCB is unlocked and most 5796 * likely gone, return value of 0, the TCP is still 5797 * locked. 5798 */ 5799 static int 5800 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 5801 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5802 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5803 { 5804 int32_t ret_val = 0; 5805 5806 rack_calc_rwin(so, tp); 5807 if (thflags & TH_RST) 5808 return (rack_process_rst(m, th, so, tp)); 5809 /* 5810 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5811 * synchronized state. 5812 */ 5813 if (thflags & TH_SYN) { 5814 rack_challenge_ack(m, th, tp, &ret_val); 5815 return (ret_val); 5816 } 5817 /* 5818 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5819 * it's less than ts_recent, drop it. 5820 */ 5821 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5822 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5823 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5824 return (ret_val); 5825 } 5826 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5827 return (ret_val); 5828 } 5829 /* 5830 * If last ACK falls within this segment's sequence numbers, record 5831 * its timestamp. NOTE: 1) That the test incorporates suggestions 5832 * from the latest proposal of the tcplw@cray.com list (Braden 5833 * 1993/04/26). 2) That updating only on newer timestamps interferes 5834 * with our earlier PAWS tests, so this check should be solely 5835 * predicated on the sequence space of this segment. 3) That we 5836 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5837 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5838 * SEG.Len, This modified check allows us to overcome RFC1323's 5839 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5840 * p.869. In such cases, we can still calculate the RTT correctly 5841 * when RCV.NXT == Last.ACK.Sent. 5842 */ 5843 if ((to->to_flags & TOF_TS) != 0 && 5844 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5845 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5846 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5847 tp->ts_recent_age = tcp_ts_getticks(); 5848 tp->ts_recent = to->to_tsval; 5849 } 5850 /* 5851 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5852 * is on (half-synchronized state), then queue data for later 5853 * processing; else drop segment and return. 5854 */ 5855 if ((thflags & TH_ACK) == 0) { 5856 if (tp->t_flags & TF_NEEDSYN) { 5857 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5858 tiwin, thflags, nxt_pkt)); 5859 5860 } else if (tp->t_flags & TF_ACKNOW) { 5861 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5862 return (ret_val); 5863 } else { 5864 rack_do_drop(m, NULL); 5865 return (0); 5866 } 5867 } 5868 /* 5869 * Ack processing. 5870 */ 5871 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 5872 return (ret_val); 5873 } 5874 if (sbavail(&so->so_snd)) { 5875 if (rack_progress_timeout_check(tp)) { 5876 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 5877 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 5878 return (1); 5879 } 5880 } 5881 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5882 tiwin, thflags, nxt_pkt)); 5883 } 5884 5885 static int 5886 rack_check_data_after_close(struct mbuf *m, 5887 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) 5888 { 5889 struct tcp_rack *rack; 5890 5891 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 5892 rack = (struct tcp_rack *)tp->t_fb_ptr; 5893 if (rack->rc_allow_data_af_clo == 0) { 5894 close_now: 5895 tp = tcp_close(tp); 5896 TCPSTAT_INC(tcps_rcvafterclose); 5897 rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 5898 return (1); 5899 } 5900 if (sbavail(&so->so_snd) == 0) 5901 goto close_now; 5902 /* Ok we allow data that is ignored and a followup reset */ 5903 tp->rcv_nxt = th->th_seq + *tlen; 5904 tp->t_flags2 |= TF2_DROP_AF_DATA; 5905 rack->r_wanted_output = 1; 5906 *tlen = 0; 5907 return (0); 5908 } 5909 5910 /* 5911 * Return value of 1, the TCB is unlocked and most 5912 * likely gone, return value of 0, the TCP is still 5913 * locked. 5914 */ 5915 static int 5916 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 5917 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 5918 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 5919 { 5920 int32_t ret_val = 0; 5921 int32_t ourfinisacked = 0; 5922 5923 rack_calc_rwin(so, tp); 5924 5925 if (thflags & TH_RST) 5926 return (rack_process_rst(m, th, so, tp)); 5927 /* 5928 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 5929 * synchronized state. 5930 */ 5931 if (thflags & TH_SYN) { 5932 rack_challenge_ack(m, th, tp, &ret_val); 5933 return (ret_val); 5934 } 5935 /* 5936 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 5937 * it's less than ts_recent, drop it. 5938 */ 5939 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 5940 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 5941 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 5942 return (ret_val); 5943 } 5944 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 5945 return (ret_val); 5946 } 5947 /* 5948 * If new data are received on a connection after the user processes 5949 * are gone, then RST the other end. 5950 */ 5951 if ((so->so_state & SS_NOFDREF) && tlen) { 5952 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 5953 return (1); 5954 } 5955 /* 5956 * If last ACK falls within this segment's sequence numbers, record 5957 * its timestamp. NOTE: 1) That the test incorporates suggestions 5958 * from the latest proposal of the tcplw@cray.com list (Braden 5959 * 1993/04/26). 2) That updating only on newer timestamps interferes 5960 * with our earlier PAWS tests, so this check should be solely 5961 * predicated on the sequence space of this segment. 3) That we 5962 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 5963 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 5964 * SEG.Len, This modified check allows us to overcome RFC1323's 5965 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 5966 * p.869. In such cases, we can still calculate the RTT correctly 5967 * when RCV.NXT == Last.ACK.Sent. 5968 */ 5969 if ((to->to_flags & TOF_TS) != 0 && 5970 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 5971 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 5972 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 5973 tp->ts_recent_age = tcp_ts_getticks(); 5974 tp->ts_recent = to->to_tsval; 5975 } 5976 /* 5977 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 5978 * is on (half-synchronized state), then queue data for later 5979 * processing; else drop segment and return. 5980 */ 5981 if ((thflags & TH_ACK) == 0) { 5982 if (tp->t_flags & TF_NEEDSYN) { 5983 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 5984 tiwin, thflags, nxt_pkt)); 5985 } else if (tp->t_flags & TF_ACKNOW) { 5986 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 5987 return (ret_val); 5988 } else { 5989 rack_do_drop(m, NULL); 5990 return (0); 5991 } 5992 } 5993 /* 5994 * Ack processing. 5995 */ 5996 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 5997 return (ret_val); 5998 } 5999 if (ourfinisacked) { 6000 /* 6001 * If we can't receive any more data, then closing user can 6002 * proceed. Starting the timer is contrary to the 6003 * specification, but if we don't get a FIN we'll hang 6004 * forever. 6005 * 6006 * XXXjl: we should release the tp also, and use a 6007 * compressed state. 6008 */ 6009 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 6010 soisdisconnected(so); 6011 tcp_timer_activate(tp, TT_2MSL, 6012 (tcp_fast_finwait2_recycle ? 6013 tcp_finwait2_timeout : 6014 TP_MAXIDLE(tp))); 6015 } 6016 tcp_state_change(tp, TCPS_FIN_WAIT_2); 6017 } 6018 if (sbavail(&so->so_snd)) { 6019 if (rack_progress_timeout_check(tp)) { 6020 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6021 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6022 return (1); 6023 } 6024 } 6025 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6026 tiwin, thflags, nxt_pkt)); 6027 } 6028 6029 /* 6030 * Return value of 1, the TCB is unlocked and most 6031 * likely gone, return value of 0, the TCP is still 6032 * locked. 6033 */ 6034 static int 6035 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 6036 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6037 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6038 { 6039 int32_t ret_val = 0; 6040 int32_t ourfinisacked = 0; 6041 6042 rack_calc_rwin(so, tp); 6043 6044 if (thflags & TH_RST) 6045 return (rack_process_rst(m, th, so, tp)); 6046 /* 6047 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6048 * synchronized state. 6049 */ 6050 if (thflags & TH_SYN) { 6051 rack_challenge_ack(m, th, tp, &ret_val); 6052 return (ret_val); 6053 } 6054 /* 6055 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6056 * it's less than ts_recent, drop it. 6057 */ 6058 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6059 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6060 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6061 return (ret_val); 6062 } 6063 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6064 return (ret_val); 6065 } 6066 /* 6067 * If new data are received on a connection after the user processes 6068 * are gone, then RST the other end. 6069 */ 6070 if ((so->so_state & SS_NOFDREF) && tlen) { 6071 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6072 return (1); 6073 } 6074 /* 6075 * If last ACK falls within this segment's sequence numbers, record 6076 * its timestamp. NOTE: 1) That the test incorporates suggestions 6077 * from the latest proposal of the tcplw@cray.com list (Braden 6078 * 1993/04/26). 2) That updating only on newer timestamps interferes 6079 * with our earlier PAWS tests, so this check should be solely 6080 * predicated on the sequence space of this segment. 3) That we 6081 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6082 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6083 * SEG.Len, This modified check allows us to overcome RFC1323's 6084 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6085 * p.869. In such cases, we can still calculate the RTT correctly 6086 * when RCV.NXT == Last.ACK.Sent. 6087 */ 6088 if ((to->to_flags & TOF_TS) != 0 && 6089 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6090 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6091 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6092 tp->ts_recent_age = tcp_ts_getticks(); 6093 tp->ts_recent = to->to_tsval; 6094 } 6095 /* 6096 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6097 * is on (half-synchronized state), then queue data for later 6098 * processing; else drop segment and return. 6099 */ 6100 if ((thflags & TH_ACK) == 0) { 6101 if (tp->t_flags & TF_NEEDSYN) { 6102 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6103 tiwin, thflags, nxt_pkt)); 6104 } else if (tp->t_flags & TF_ACKNOW) { 6105 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6106 return (ret_val); 6107 } else { 6108 rack_do_drop(m, NULL); 6109 return (0); 6110 } 6111 } 6112 /* 6113 * Ack processing. 6114 */ 6115 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6116 return (ret_val); 6117 } 6118 if (ourfinisacked) { 6119 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6120 tcp_twstart(tp); 6121 m_freem(m); 6122 return (1); 6123 } 6124 if (sbavail(&so->so_snd)) { 6125 if (rack_progress_timeout_check(tp)) { 6126 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6127 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6128 return (1); 6129 } 6130 } 6131 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6132 tiwin, thflags, nxt_pkt)); 6133 } 6134 6135 /* 6136 * Return value of 1, the TCB is unlocked and most 6137 * likely gone, return value of 0, the TCP is still 6138 * locked. 6139 */ 6140 static int 6141 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 6142 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6143 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6144 { 6145 int32_t ret_val = 0; 6146 int32_t ourfinisacked = 0; 6147 6148 rack_calc_rwin(so, tp); 6149 6150 if (thflags & TH_RST) 6151 return (rack_process_rst(m, th, so, tp)); 6152 /* 6153 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6154 * synchronized state. 6155 */ 6156 if (thflags & TH_SYN) { 6157 rack_challenge_ack(m, th, tp, &ret_val); 6158 return (ret_val); 6159 } 6160 /* 6161 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6162 * it's less than ts_recent, drop it. 6163 */ 6164 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6165 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6166 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6167 return (ret_val); 6168 } 6169 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6170 return (ret_val); 6171 } 6172 /* 6173 * If new data are received on a connection after the user processes 6174 * are gone, then RST the other end. 6175 */ 6176 if ((so->so_state & SS_NOFDREF) && tlen) { 6177 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6178 return (1); 6179 } 6180 /* 6181 * If last ACK falls within this segment's sequence numbers, record 6182 * its timestamp. NOTE: 1) That the test incorporates suggestions 6183 * from the latest proposal of the tcplw@cray.com list (Braden 6184 * 1993/04/26). 2) That updating only on newer timestamps interferes 6185 * with our earlier PAWS tests, so this check should be solely 6186 * predicated on the sequence space of this segment. 3) That we 6187 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6188 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6189 * SEG.Len, This modified check allows us to overcome RFC1323's 6190 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6191 * p.869. In such cases, we can still calculate the RTT correctly 6192 * when RCV.NXT == Last.ACK.Sent. 6193 */ 6194 if ((to->to_flags & TOF_TS) != 0 && 6195 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6196 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6197 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6198 tp->ts_recent_age = tcp_ts_getticks(); 6199 tp->ts_recent = to->to_tsval; 6200 } 6201 /* 6202 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6203 * is on (half-synchronized state), then queue data for later 6204 * processing; else drop segment and return. 6205 */ 6206 if ((thflags & TH_ACK) == 0) { 6207 if (tp->t_flags & TF_NEEDSYN) { 6208 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6209 tiwin, thflags, nxt_pkt)); 6210 } else if (tp->t_flags & TF_ACKNOW) { 6211 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6212 return (ret_val); 6213 } else { 6214 rack_do_drop(m, NULL); 6215 return (0); 6216 } 6217 } 6218 /* 6219 * case TCPS_LAST_ACK: Ack processing. 6220 */ 6221 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6222 return (ret_val); 6223 } 6224 if (ourfinisacked) { 6225 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); 6226 tp = tcp_close(tp); 6227 rack_do_drop(m, tp); 6228 return (1); 6229 } 6230 if (sbavail(&so->so_snd)) { 6231 if (rack_progress_timeout_check(tp)) { 6232 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6233 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6234 return (1); 6235 } 6236 } 6237 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6238 tiwin, thflags, nxt_pkt)); 6239 } 6240 6241 6242 /* 6243 * Return value of 1, the TCB is unlocked and most 6244 * likely gone, return value of 0, the TCP is still 6245 * locked. 6246 */ 6247 static int 6248 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 6249 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 6250 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 6251 { 6252 int32_t ret_val = 0; 6253 int32_t ourfinisacked = 0; 6254 6255 rack_calc_rwin(so, tp); 6256 6257 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 6258 if (thflags & TH_RST) 6259 return (rack_process_rst(m, th, so, tp)); 6260 /* 6261 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 6262 * synchronized state. 6263 */ 6264 if (thflags & TH_SYN) { 6265 rack_challenge_ack(m, th, tp, &ret_val); 6266 return (ret_val); 6267 } 6268 /* 6269 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 6270 * it's less than ts_recent, drop it. 6271 */ 6272 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 6273 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 6274 if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val)) 6275 return (ret_val); 6276 } 6277 if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 6278 return (ret_val); 6279 } 6280 /* 6281 * If new data are received on a connection after the user processes 6282 * are gone, then RST the other end. 6283 */ 6284 if ((so->so_state & SS_NOFDREF) && 6285 tlen) { 6286 if (rack_check_data_after_close(m, tp, &tlen, th, so)) 6287 return (1); 6288 } 6289 /* 6290 * If last ACK falls within this segment's sequence numbers, record 6291 * its timestamp. NOTE: 1) That the test incorporates suggestions 6292 * from the latest proposal of the tcplw@cray.com list (Braden 6293 * 1993/04/26). 2) That updating only on newer timestamps interferes 6294 * with our earlier PAWS tests, so this check should be solely 6295 * predicated on the sequence space of this segment. 3) That we 6296 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 6297 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 6298 * SEG.Len, This modified check allows us to overcome RFC1323's 6299 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 6300 * p.869. In such cases, we can still calculate the RTT correctly 6301 * when RCV.NXT == Last.ACK.Sent. 6302 */ 6303 if ((to->to_flags & TOF_TS) != 0 && 6304 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 6305 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 6306 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 6307 tp->ts_recent_age = tcp_ts_getticks(); 6308 tp->ts_recent = to->to_tsval; 6309 } 6310 /* 6311 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 6312 * is on (half-synchronized state), then queue data for later 6313 * processing; else drop segment and return. 6314 */ 6315 if ((thflags & TH_ACK) == 0) { 6316 if (tp->t_flags & TF_NEEDSYN) { 6317 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6318 tiwin, thflags, nxt_pkt)); 6319 } else if (tp->t_flags & TF_ACKNOW) { 6320 rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 6321 return (ret_val); 6322 } else { 6323 rack_do_drop(m, NULL); 6324 return (0); 6325 } 6326 } 6327 /* 6328 * Ack processing. 6329 */ 6330 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 6331 return (ret_val); 6332 } 6333 if (sbavail(&so->so_snd)) { 6334 if (rack_progress_timeout_check(tp)) { 6335 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 6336 rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 6337 return (1); 6338 } 6339 } 6340 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, 6341 tiwin, thflags, nxt_pkt)); 6342 } 6343 6344 6345 static void inline 6346 rack_clear_rate_sample(struct tcp_rack *rack) 6347 { 6348 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; 6349 rack->r_ctl.rack_rs.rs_rtt_cnt = 0; 6350 rack->r_ctl.rack_rs.rs_rtt_tot = 0; 6351 } 6352 6353 static int 6354 rack_init(struct tcpcb *tp) 6355 { 6356 struct tcp_rack *rack = NULL; 6357 6358 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); 6359 if (tp->t_fb_ptr == NULL) { 6360 /* 6361 * We need to allocate memory but cant. The INP and INP_INFO 6362 * locks and they are recusive (happens during setup. So a 6363 * scheme to drop the locks fails :( 6364 * 6365 */ 6366 return (ENOMEM); 6367 } 6368 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); 6369 6370 rack = (struct tcp_rack *)tp->t_fb_ptr; 6371 TAILQ_INIT(&rack->r_ctl.rc_map); 6372 TAILQ_INIT(&rack->r_ctl.rc_free); 6373 TAILQ_INIT(&rack->r_ctl.rc_tmap); 6374 rack->rc_tp = tp; 6375 if (tp->t_inpcb) { 6376 rack->rc_inp = tp->t_inpcb; 6377 } 6378 /* Probably not needed but lets be sure */ 6379 rack_clear_rate_sample(rack); 6380 rack->r_cpu = 0; 6381 rack->r_ctl.rc_reorder_fade = rack_reorder_fade; 6382 rack->rc_allow_data_af_clo = rack_ignore_data_after_close; 6383 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; 6384 rack->rc_pace_reduce = rack_slot_reduction; 6385 if (V_tcp_delack_enabled) 6386 tp->t_delayed_ack = 1; 6387 else 6388 tp->t_delayed_ack = 0; 6389 rack->rc_pace_max_segs = rack_hptsi_segments; 6390 rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; 6391 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; 6392 rack->r_ctl.rc_pkt_delay = rack_pkt_delay; 6393 rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; 6394 rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; 6395 rack->r_enforce_min_pace = rack_min_pace_time; 6396 rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; 6397 rack->r_ctl.rc_prop_rate = rack_proportional_rate; 6398 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; 6399 rack->r_ctl.rc_early_recovery = rack_early_recovery; 6400 rack->rc_always_pace = rack_pace_every_seg; 6401 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; 6402 rack->rack_tlp_threshold_use = rack_tlp_threshold_use; 6403 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; 6404 rack->r_ctl.rc_min_to = rack_min_to; 6405 rack->r_ctl.rc_prr_inc_var = rack_inc_var; 6406 if (tp->snd_una != tp->snd_max) { 6407 /* Create a send map for the current outstanding data */ 6408 struct rack_sendmap *rsm; 6409 6410 rsm = rack_alloc(rack); 6411 if (rsm == NULL) { 6412 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6413 tp->t_fb_ptr = NULL; 6414 return (ENOMEM); 6415 } 6416 rsm->r_flags = RACK_OVERMAX; 6417 rsm->r_tim_lastsent[0] = tcp_ts_getticks(); 6418 rsm->r_rtr_cnt = 1; 6419 rsm->r_rtr_bytes = 0; 6420 rsm->r_start = tp->snd_una; 6421 rsm->r_end = tp->snd_max; 6422 rsm->r_sndcnt = 0; 6423 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); 6424 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); 6425 rsm->r_in_tmap = 1; 6426 } 6427 rack_stop_all_timers(tp); 6428 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6429 return (0); 6430 } 6431 6432 static int 6433 rack_handoff_ok(struct tcpcb *tp) 6434 { 6435 if ((tp->t_state == TCPS_CLOSED) || 6436 (tp->t_state == TCPS_LISTEN)) { 6437 /* Sure no problem though it may not stick */ 6438 return (0); 6439 } 6440 if ((tp->t_state == TCPS_SYN_SENT) || 6441 (tp->t_state == TCPS_SYN_RECEIVED)) { 6442 /* 6443 * We really don't know you have to get to ESTAB or beyond 6444 * to tell. 6445 */ 6446 return (EAGAIN); 6447 } 6448 if (tp->t_flags & TF_SACK_PERMIT) { 6449 return (0); 6450 } 6451 /* 6452 * If we reach here we don't do SACK on this connection so we can 6453 * never do rack. 6454 */ 6455 return (EINVAL); 6456 } 6457 6458 static void 6459 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) 6460 { 6461 if (tp->t_fb_ptr) { 6462 struct tcp_rack *rack; 6463 struct rack_sendmap *rsm; 6464 6465 rack = (struct tcp_rack *)tp->t_fb_ptr; 6466 #ifdef TCP_BLACKBOX 6467 tcp_log_flowend(tp); 6468 #endif 6469 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6470 while (rsm) { 6471 TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); 6472 uma_zfree(rack_zone, rsm); 6473 rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); 6474 } 6475 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6476 while (rsm) { 6477 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); 6478 uma_zfree(rack_zone, rsm); 6479 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); 6480 } 6481 rack->rc_free_cnt = 0; 6482 uma_zfree(rack_pcb_zone, tp->t_fb_ptr); 6483 tp->t_fb_ptr = NULL; 6484 } 6485 /* Make sure snd_nxt is correctly set */ 6486 tp->snd_nxt = tp->snd_max; 6487 } 6488 6489 static void 6490 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) 6491 { 6492 switch (tp->t_state) { 6493 case TCPS_SYN_SENT: 6494 rack->r_state = TCPS_SYN_SENT; 6495 rack->r_substate = rack_do_syn_sent; 6496 break; 6497 case TCPS_SYN_RECEIVED: 6498 rack->r_state = TCPS_SYN_RECEIVED; 6499 rack->r_substate = rack_do_syn_recv; 6500 break; 6501 case TCPS_ESTABLISHED: 6502 rack->r_state = TCPS_ESTABLISHED; 6503 rack->r_substate = rack_do_established; 6504 break; 6505 case TCPS_CLOSE_WAIT: 6506 rack->r_state = TCPS_CLOSE_WAIT; 6507 rack->r_substate = rack_do_close_wait; 6508 break; 6509 case TCPS_FIN_WAIT_1: 6510 rack->r_state = TCPS_FIN_WAIT_1; 6511 rack->r_substate = rack_do_fin_wait_1; 6512 break; 6513 case TCPS_CLOSING: 6514 rack->r_state = TCPS_CLOSING; 6515 rack->r_substate = rack_do_closing; 6516 break; 6517 case TCPS_LAST_ACK: 6518 rack->r_state = TCPS_LAST_ACK; 6519 rack->r_substate = rack_do_lastack; 6520 break; 6521 case TCPS_FIN_WAIT_2: 6522 rack->r_state = TCPS_FIN_WAIT_2; 6523 rack->r_substate = rack_do_fin_wait_2; 6524 break; 6525 case TCPS_LISTEN: 6526 case TCPS_CLOSED: 6527 case TCPS_TIME_WAIT: 6528 default: 6529 break; 6530 }; 6531 } 6532 6533 6534 static void 6535 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) 6536 { 6537 /* 6538 * We received an ack, and then did not 6539 * call send or were bounced out due to the 6540 * hpts was running. Now a timer is up as well, is 6541 * it the right timer? 6542 */ 6543 struct rack_sendmap *rsm; 6544 int tmr_up; 6545 6546 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 6547 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 6548 return; 6549 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6550 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 6551 (tmr_up == PACE_TMR_RXT)) { 6552 /* Should be an RXT */ 6553 return; 6554 } 6555 if (rsm == NULL) { 6556 /* Nothing outstanding? */ 6557 if (tp->t_flags & TF_DELACK) { 6558 if (tmr_up == PACE_TMR_DELACK) 6559 /* We are supposed to have delayed ack up and we do */ 6560 return; 6561 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { 6562 /* 6563 * if we hit enobufs then we would expect the possiblity 6564 * of nothing outstanding and the RXT up (and the hptsi timer). 6565 */ 6566 return; 6567 } else if (((tcp_always_keepalive || 6568 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6569 (tp->t_state <= TCPS_CLOSING)) && 6570 (tmr_up == PACE_TMR_KEEP) && 6571 (tp->snd_max == tp->snd_una)) { 6572 /* We should have keep alive up and we do */ 6573 return; 6574 } 6575 } 6576 if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { 6577 if ((tp->t_flags & TF_SENTFIN) && 6578 ((tp->snd_max - tp->snd_una) == 1) && 6579 (rsm->r_flags & RACK_HAS_FIN)) { 6580 /* needs to be a RXT */ 6581 if (tmr_up == PACE_TMR_RXT) 6582 return; 6583 } else if (tmr_up == PACE_TMR_RACK) 6584 return; 6585 } else if (SEQ_GT(tp->snd_max,tp->snd_una) && 6586 ((tmr_up == PACE_TMR_TLP) || 6587 (tmr_up == PACE_TMR_RXT))) { 6588 /* 6589 * Either a TLP or RXT is fine if no sack-passed 6590 * is in place and data is outstanding. 6591 */ 6592 return; 6593 } else if (tmr_up == PACE_TMR_DELACK) { 6594 /* 6595 * If the delayed ack was going to go off 6596 * before the rtx/tlp/rack timer were going to 6597 * expire, then that would be the timer in control. 6598 * Note we don't check the time here trusting the 6599 * code is correct. 6600 */ 6601 return; 6602 } 6603 /* 6604 * Ok the timer originally started is not what we want now. 6605 * We will force the hpts to be stopped if any, and restart 6606 * with the slot set to what was in the saved slot. 6607 */ 6608 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 6609 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6610 } 6611 6612 static void 6613 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6614 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 6615 int32_t nxt_pkt, struct timeval *tv) 6616 { 6617 int32_t thflags, retval, did_out = 0; 6618 int32_t way_out = 0; 6619 uint32_t cts; 6620 uint32_t tiwin; 6621 struct tcpopt to; 6622 struct tcp_rack *rack; 6623 struct rack_sendmap *rsm; 6624 int32_t prev_state = 0; 6625 6626 cts = tcp_tv_to_mssectick(tv); 6627 rack = (struct tcp_rack *)tp->t_fb_ptr; 6628 6629 kern_prefetch(rack, &prev_state); 6630 prev_state = 0; 6631 thflags = th->th_flags; 6632 /* 6633 * If this is either a state-changing packet or current state isn't 6634 * established, we require a read lock on tcbinfo. Otherwise, we 6635 * allow the tcbinfo to be in either locked or unlocked, as the 6636 * caller may have unnecessarily acquired a lock due to a race. 6637 */ 6638 INP_WLOCK_ASSERT(tp->t_inpcb); 6639 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 6640 __func__)); 6641 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 6642 __func__)); 6643 { 6644 union tcp_log_stackspecific log; 6645 6646 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 6647 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 6648 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 6649 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; 6650 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, 6651 tlen, &log, true); 6652 } 6653 /* 6654 * Segment received on connection. Reset idle time and keep-alive 6655 * timer. XXX: This should be done after segment validation to 6656 * ignore broken/spoofed segs. 6657 */ 6658 if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { 6659 if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { 6660 counter_u64_add(rack_input_idle_reduces, 1); 6661 rack_cc_after_idle(tp, 6662 (rack->r_idle_reduce_largest ? 1 :0)); 6663 } 6664 } 6665 rack->r_ctl.rc_rcvtime = cts; 6666 tp->t_rcvtime = ticks; 6667 6668 /* 6669 * Unscale the window into a 32-bit value. For the SYN_SENT state 6670 * the scale is zero. 6671 */ 6672 tiwin = th->th_win << tp->snd_scale; 6673 #ifdef NETFLIX_STATS 6674 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 6675 #endif 6676 /* 6677 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move 6678 * this to occur after we've validated the segment. 6679 */ 6680 if (tp->t_flags & TF_ECN_PERMIT) { 6681 if (thflags & TH_CWR) 6682 tp->t_flags &= ~TF_ECN_SND_ECE; 6683 switch (iptos & IPTOS_ECN_MASK) { 6684 case IPTOS_ECN_CE: 6685 tp->t_flags |= TF_ECN_SND_ECE; 6686 TCPSTAT_INC(tcps_ecn_ce); 6687 break; 6688 case IPTOS_ECN_ECT0: 6689 TCPSTAT_INC(tcps_ecn_ect0); 6690 break; 6691 case IPTOS_ECN_ECT1: 6692 TCPSTAT_INC(tcps_ecn_ect1); 6693 break; 6694 } 6695 /* Congestion experienced. */ 6696 if (thflags & TH_ECE) { 6697 rack_cong_signal(tp, th, CC_ECN); 6698 } 6699 } 6700 /* 6701 * Parse options on any incoming segment. 6702 */ 6703 tcp_dooptions(&to, (u_char *)(th + 1), 6704 (th->th_off << 2) - sizeof(struct tcphdr), 6705 (thflags & TH_SYN) ? TO_SYN : 0); 6706 6707 /* 6708 * If echoed timestamp is later than the current time, fall back to 6709 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 6710 * were used when this connection was established. 6711 */ 6712 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 6713 to.to_tsecr -= tp->ts_offset; 6714 if (TSTMP_GT(to.to_tsecr, cts)) 6715 to.to_tsecr = 0; 6716 } 6717 /* 6718 * If its the first time in we need to take care of options and 6719 * verify we can do SACK for rack! 6720 */ 6721 if (rack->r_state == 0) { 6722 /* Should be init'd by rack_init() */ 6723 KASSERT(rack->rc_inp != NULL, 6724 ("%s: rack->rc_inp unexpectedly NULL", __func__)); 6725 if (rack->rc_inp == NULL) { 6726 rack->rc_inp = tp->t_inpcb; 6727 } 6728 6729 /* 6730 * Process options only when we get SYN/ACK back. The SYN 6731 * case for incoming connections is handled in tcp_syncache. 6732 * According to RFC1323 the window field in a SYN (i.e., a 6733 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 6734 * this is traditional behavior, may need to be cleaned up. 6735 */ 6736 rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 6737 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 6738 if ((to.to_flags & TOF_SCALE) && 6739 (tp->t_flags & TF_REQ_SCALE)) { 6740 tp->t_flags |= TF_RCVD_SCALE; 6741 tp->snd_scale = to.to_wscale; 6742 } 6743 /* 6744 * Initial send window. It will be updated with the 6745 * next incoming segment to the scaled value. 6746 */ 6747 tp->snd_wnd = th->th_win; 6748 if (to.to_flags & TOF_TS) { 6749 tp->t_flags |= TF_RCVD_TSTMP; 6750 tp->ts_recent = to.to_tsval; 6751 tp->ts_recent_age = cts; 6752 } 6753 if (to.to_flags & TOF_MSS) 6754 tcp_mss(tp, to.to_mss); 6755 if ((tp->t_flags & TF_SACK_PERMIT) && 6756 (to.to_flags & TOF_SACKPERM) == 0) 6757 tp->t_flags &= ~TF_SACK_PERMIT; 6758 } 6759 /* 6760 * At this point we are at the initial call. Here we decide 6761 * if we are doing RACK or not. We do this by seeing if 6762 * TF_SACK_PERMIT is set, if not rack is *not* possible and 6763 * we switch to the default code. 6764 */ 6765 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 6766 tcp_switch_back_to_default(tp); 6767 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 6768 tlen, iptos); 6769 return; 6770 } 6771 /* Set the flag */ 6772 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 6773 tcp_set_hpts(tp->t_inpcb); 6774 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); 6775 } 6776 /* 6777 * This is the one exception case where we set the rack state 6778 * always. All other times (timers etc) we must have a rack-state 6779 * set (so we assure we have done the checks above for SACK). 6780 */ 6781 if (rack->r_state != tp->t_state) 6782 rack_set_state(tp, rack); 6783 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) 6784 kern_prefetch(rsm, &prev_state); 6785 prev_state = rack->r_state; 6786 rack->r_ctl.rc_tlp_send_cnt = 0; 6787 rack_clear_rate_sample(rack); 6788 retval = (*rack->r_substate) (m, th, so, 6789 tp, &to, drop_hdrlen, 6790 tlen, tiwin, thflags, nxt_pkt); 6791 #ifdef INVARIANTS 6792 if ((retval == 0) && 6793 (tp->t_inpcb == NULL)) { 6794 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 6795 retval, tp, prev_state); 6796 } 6797 #endif 6798 if (retval == 0) { 6799 /* 6800 * If retval is 1 the tcb is unlocked and most likely the tp 6801 * is gone. 6802 */ 6803 INP_WLOCK_ASSERT(tp->t_inpcb); 6804 tcp_rack_xmit_timer_commit(rack, tp); 6805 if (nxt_pkt == 0) { 6806 if (rack->r_wanted_output != 0) { 6807 did_out = 1; 6808 (void)tp->t_fb->tfb_tcp_output(tp); 6809 } 6810 rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); 6811 } 6812 if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 6813 (SEQ_GT(tp->snd_max, tp->snd_una) || 6814 (tp->t_flags & TF_DELACK) || 6815 ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 6816 (tp->t_state <= TCPS_CLOSING)))) { 6817 /* We could not send (probably in the hpts but stopped the timer earlier)? */ 6818 if ((tp->snd_max == tp->snd_una) && 6819 ((tp->t_flags & TF_DELACK) == 0) && 6820 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 6821 /* keep alive not needed if we are hptsi output yet */ 6822 ; 6823 } else { 6824 if (rack->rc_inp->inp_in_hpts) 6825 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); 6826 rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); 6827 } 6828 way_out = 1; 6829 } else { 6830 /* Do we have the correct timer running? */ 6831 rack_timer_audit(tp, rack, &so->so_snd); 6832 way_out = 2; 6833 } 6834 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); 6835 if (did_out) 6836 rack->r_wanted_output = 0; 6837 #ifdef INVARIANTS 6838 if (tp->t_inpcb == NULL) { 6839 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 6840 did_out, 6841 retval, tp, prev_state); 6842 } 6843 #endif 6844 INP_WUNLOCK(tp->t_inpcb); 6845 } 6846 } 6847 6848 void 6849 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 6850 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 6851 { 6852 struct timeval tv; 6853 #ifdef RSS 6854 struct tcp_function_block *tfb; 6855 struct tcp_rack *rack; 6856 struct inpcb *inp; 6857 6858 rack = (struct tcp_rack *)tp->t_fb_ptr; 6859 if (rack->r_state == 0) { 6860 /* 6861 * Initial input (ACK to SYN-ACK etc)lets go ahead and get 6862 * it processed 6863 */ 6864 tcp_get_usecs(&tv); 6865 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6866 tlen, iptos, 0, &tv); 6867 return; 6868 } 6869 tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos); 6870 INP_WUNLOCK(tp->t_inpcb); 6871 #else 6872 tcp_get_usecs(&tv); 6873 rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, 6874 tlen, iptos, 0, &tv); 6875 #endif 6876 } 6877 6878 struct rack_sendmap * 6879 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) 6880 { 6881 struct rack_sendmap *rsm = NULL; 6882 int32_t idx; 6883 uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; 6884 6885 /* Return the next guy to be re-transmitted */ 6886 if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { 6887 return (NULL); 6888 } 6889 if (tp->t_flags & TF_SENTFIN) { 6890 /* retran the end FIN? */ 6891 return (NULL); 6892 } 6893 /* ok lets look at this one */ 6894 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); 6895 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { 6896 goto check_it; 6897 } 6898 rsm = rack_find_lowest_rsm(rack); 6899 if (rsm == NULL) { 6900 return (NULL); 6901 } 6902 check_it: 6903 srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; 6904 srtt = TICKS_2_MSEC(srtt_cur); 6905 if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) 6906 srtt = rack->rc_rack_rtt; 6907 if (rsm->r_flags & RACK_ACKED) { 6908 return (NULL); 6909 } 6910 if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { 6911 /* Its not yet ready */ 6912 return (NULL); 6913 } 6914 idx = rsm->r_rtr_cnt - 1; 6915 ts_low = rsm->r_tim_lastsent[idx]; 6916 thresh = rack_calc_thresh_rack(rack, srtt, tsused); 6917 if (tsused <= ts_low) { 6918 return (NULL); 6919 } 6920 if ((tsused - ts_low) >= thresh) { 6921 return (rsm); 6922 } 6923 return (NULL); 6924 } 6925 6926 static int 6927 rack_output(struct tcpcb *tp) 6928 { 6929 struct socket *so; 6930 uint32_t recwin, sendwin; 6931 uint32_t sb_offset; 6932 int32_t len, flags, error = 0; 6933 struct mbuf *m; 6934 struct mbuf *mb; 6935 uint32_t if_hw_tsomaxsegcount = 0; 6936 uint32_t if_hw_tsomaxsegsize; 6937 long tot_len_this_send = 0; 6938 struct ip *ip = NULL; 6939 #ifdef TCPDEBUG 6940 struct ipovly *ipov = NULL; 6941 #endif 6942 #ifdef NETFLIX_TCP_O_UDP 6943 struct udphdr *udp = NULL; 6944 #endif 6945 struct tcp_rack *rack; 6946 struct tcphdr *th; 6947 uint8_t pass = 0; 6948 u_char opt[TCP_MAXOLEN]; 6949 unsigned ipoptlen, optlen, hdrlen; 6950 #ifdef NETFLIX_TCP_O_UDP 6951 unsigned ulen; 6952 #endif 6953 uint32_t rack_seq; 6954 6955 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 6956 unsigned ipsec_optlen = 0; 6957 6958 #endif 6959 int32_t idle, sendalot; 6960 int32_t sub_from_prr = 0; 6961 volatile int32_t sack_rxmit; 6962 struct rack_sendmap *rsm = NULL; 6963 int32_t tso, mtu, would_have_fin = 0; 6964 struct tcpopt to; 6965 int32_t slot = 0; 6966 uint32_t cts; 6967 uint8_t hpts_calling, doing_tlp = 0; 6968 int32_t do_a_prefetch; 6969 int32_t prefetch_rsm = 0; 6970 int32_t prefetch_so_done = 0; 6971 struct tcp_log_buffer *lgb = NULL; 6972 struct inpcb *inp; 6973 struct sockbuf *sb; 6974 #ifdef INET6 6975 struct ip6_hdr *ip6 = NULL; 6976 int32_t isipv6; 6977 #endif 6978 /* setup and take the cache hits here */ 6979 rack = (struct tcp_rack *)tp->t_fb_ptr; 6980 inp = rack->rc_inp; 6981 so = inp->inp_socket; 6982 sb = &so->so_snd; 6983 kern_prefetch(sb, &do_a_prefetch); 6984 do_a_prefetch = 1; 6985 6986 INP_WLOCK_ASSERT(inp); 6987 #ifdef TCP_OFFLOAD 6988 if (tp->t_flags & TF_TOE) 6989 return (tcp_offload_output(tp)); 6990 #endif 6991 6992 #ifdef TCP_RFC7413 6993 /* 6994 * For TFO connections in SYN_RECEIVED, only allow the initial 6995 * SYN|ACK and those sent by the retransmit timer. 6996 */ 6997 if ((tp->t_flags & TF_FASTOPEN) && 6998 (tp->t_state == TCPS_SYN_RECEIVED) && 6999 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */ 7000 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ 7001 return (0); 7002 #endif 7003 #ifdef INET6 7004 if (rack->r_state) { 7005 /* Use the cache line loaded if possible */ 7006 isipv6 = rack->r_is_v6; 7007 } else { 7008 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 7009 } 7010 #endif 7011 cts = tcp_ts_getticks(); 7012 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 7013 inp->inp_in_hpts) { 7014 /* 7015 * We are on the hpts for some timer but not hptsi output. 7016 * Remove from the hpts unconditionally. 7017 */ 7018 rack_timer_cancel(tp, rack, cts, __LINE__); 7019 } 7020 /* Mark that we have called rack_output(). */ 7021 if ((rack->r_timer_override) || 7022 (tp->t_flags & TF_FORCEDATA) || 7023 (tp->t_state < TCPS_ESTABLISHED)) { 7024 if (tp->t_inpcb->inp_in_hpts) 7025 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); 7026 } else if (tp->t_inpcb->inp_in_hpts) { 7027 /* 7028 * On the hpts you can't pass even if ACKNOW is on, we will 7029 * when the hpts fires. 7030 */ 7031 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); 7032 return (0); 7033 } 7034 hpts_calling = inp->inp_hpts_calls; 7035 inp->inp_hpts_calls = 0; 7036 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 7037 if (rack_process_timers(tp, rack, cts, hpts_calling)) { 7038 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); 7039 return (0); 7040 } 7041 } 7042 rack->r_wanted_output = 0; 7043 rack->r_timer_override = 0; 7044 /* 7045 * Determine length of data that should be transmitted, and flags 7046 * that will be used. If there is some data or critical controls 7047 * (SYN, RST) to send, then transmit; otherwise, investigate 7048 * further. 7049 */ 7050 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); 7051 if (tp->t_idle_reduce) { 7052 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) 7053 rack_cc_after_idle(tp, 7054 (rack->r_idle_reduce_largest ? 1 :0)); 7055 } 7056 tp->t_flags &= ~TF_LASTIDLE; 7057 if (idle) { 7058 if (tp->t_flags & TF_MORETOCOME) { 7059 tp->t_flags |= TF_LASTIDLE; 7060 idle = 0; 7061 } 7062 } 7063 again: 7064 /* 7065 * If we've recently taken a timeout, snd_max will be greater than 7066 * snd_nxt. There may be SACK information that allows us to avoid 7067 * resending already delivered data. Adjust snd_nxt accordingly. 7068 */ 7069 sendalot = 0; 7070 cts = tcp_ts_getticks(); 7071 tso = 0; 7072 mtu = 0; 7073 sb_offset = tp->snd_max - tp->snd_una; 7074 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 7075 7076 flags = tcp_outflags[tp->t_state]; 7077 /* 7078 * Send any SACK-generated retransmissions. If we're explicitly 7079 * trying to send out new data (when sendalot is 1), bypass this 7080 * function. If we retransmit in fast recovery mode, decrement 7081 * snd_cwnd, since we're replacing a (future) new transmission with 7082 * a retransmission now, and we previously incremented snd_cwnd in 7083 * tcp_input(). 7084 */ 7085 /* 7086 * Still in sack recovery , reset rxmit flag to zero. 7087 */ 7088 while (rack->rc_free_cnt < rack_free_cache) { 7089 rsm = rack_alloc(rack); 7090 if (rsm == NULL) { 7091 if (inp->inp_hpts_calls) 7092 /* Retry in a ms */ 7093 slot = 1; 7094 goto just_return_nolock; 7095 } 7096 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); 7097 rack->rc_free_cnt++; 7098 rsm = NULL; 7099 } 7100 if (inp->inp_hpts_calls) 7101 inp->inp_hpts_calls = 0; 7102 sack_rxmit = 0; 7103 len = 0; 7104 rsm = NULL; 7105 if (flags & TH_RST) { 7106 SOCKBUF_LOCK(sb); 7107 goto send; 7108 } 7109 if (rack->r_ctl.rc_tlpsend) { 7110 /* Tail loss probe */ 7111 long cwin; 7112 long tlen; 7113 7114 doing_tlp = 1; 7115 rsm = rack->r_ctl.rc_tlpsend; 7116 rack->r_ctl.rc_tlpsend = NULL; 7117 sack_rxmit = 1; 7118 tlen = rsm->r_end - rsm->r_start; 7119 if (tlen > tp->t_maxseg) 7120 tlen = tp->t_maxseg; 7121 #ifdef INVARIANTS 7122 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 7123 panic("tp:%p rack:%p snd_una:%u rsm:%p r_start:%u", 7124 tp, rack, tp->snd_una, rsm, rsm->r_start); 7125 } 7126 #endif 7127 sb_offset = rsm->r_start - tp->snd_una; 7128 cwin = min(tp->snd_wnd, tlen); 7129 len = cwin; 7130 } else if (rack->r_ctl.rc_resend) { 7131 /* Retransmit timer */ 7132 rsm = rack->r_ctl.rc_resend; 7133 rack->r_ctl.rc_resend = NULL; 7134 len = rsm->r_end - rsm->r_start; 7135 sack_rxmit = 1; 7136 sendalot = 0; 7137 sb_offset = rsm->r_start - tp->snd_una; 7138 if (len >= tp->t_maxseg) { 7139 len = tp->t_maxseg; 7140 } 7141 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7142 __func__, sb_offset)); 7143 } else if ((rack->rc_in_persist == 0) && 7144 ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { 7145 long tlen; 7146 7147 if ((!IN_RECOVERY(tp->t_flags)) && 7148 ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { 7149 /* Enter recovery if not induced by a time-out */ 7150 rack->r_ctl.rc_rsm_start = rsm->r_start; 7151 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; 7152 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; 7153 rack_cong_signal(tp, NULL, CC_NDUPACK); 7154 /* 7155 * When we enter recovery we need to assure we send 7156 * one packet. 7157 */ 7158 rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; 7159 } 7160 #ifdef INVARIANTS 7161 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 7162 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", 7163 tp, rack, rsm, rsm->r_start, tp->snd_una); 7164 } 7165 #endif 7166 tlen = rsm->r_end - rsm->r_start; 7167 sb_offset = rsm->r_start - tp->snd_una; 7168 if (tlen > rack->r_ctl.rc_prr_sndcnt) { 7169 len = rack->r_ctl.rc_prr_sndcnt; 7170 } else { 7171 len = tlen; 7172 } 7173 if (len >= tp->t_maxseg) { 7174 sendalot = 1; 7175 len = tp->t_maxseg; 7176 } else { 7177 sendalot = 0; 7178 if ((rack->rc_timer_up == 0) && 7179 (len < tlen)) { 7180 /* 7181 * If its not a timer don't send a partial 7182 * segment. 7183 */ 7184 len = 0; 7185 goto just_return_nolock; 7186 } 7187 } 7188 KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", 7189 __func__, sb_offset)); 7190 if (len > 0) { 7191 sub_from_prr = 1; 7192 sack_rxmit = 1; 7193 TCPSTAT_INC(tcps_sack_rexmits); 7194 TCPSTAT_ADD(tcps_sack_rexmit_bytes, 7195 min(len, tp->t_maxseg)); 7196 counter_u64_add(rack_rtm_prr_retran, 1); 7197 } 7198 } 7199 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { 7200 /* we are retransmitting the fin */ 7201 len--; 7202 if (len) { 7203 /* 7204 * When retransmitting data do *not* include the 7205 * FIN. This could happen from a TLP probe. 7206 */ 7207 flags &= ~TH_FIN; 7208 } 7209 } 7210 #ifdef INVARIANTS 7211 /* For debugging */ 7212 rack->r_ctl.rc_rsm_at_retran = rsm; 7213 #endif 7214 /* 7215 * Enforce a connection sendmap count limit if set 7216 * as long as we are not retransmiting. 7217 */ 7218 if ((rsm == NULL) && 7219 (rack_map_entries_limit > 0) && 7220 (rack->r_ctl.rc_num_maps_alloced >= rack_map_entries_limit)) { 7221 counter_u64_add(rack_to_alloc_limited, 1); 7222 if (!rack->alloc_limit_reported) { 7223 rack->alloc_limit_reported = 1; 7224 counter_u64_add(rack_alloc_limited_conns, 1); 7225 } 7226 goto just_return_nolock; 7227 } 7228 /* 7229 * Get standard flags, and add SYN or FIN if requested by 'hidden' 7230 * state flags. 7231 */ 7232 if (tp->t_flags & TF_NEEDFIN) 7233 flags |= TH_FIN; 7234 if (tp->t_flags & TF_NEEDSYN) 7235 flags |= TH_SYN; 7236 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 7237 void *end_rsm; 7238 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); 7239 if (end_rsm) 7240 kern_prefetch(end_rsm, &prefetch_rsm); 7241 prefetch_rsm = 1; 7242 } 7243 SOCKBUF_LOCK(sb); 7244 /* 7245 * If in persist timeout with window of 0, send 1 byte. Otherwise, 7246 * if window is small but nonzero and time TF_SENTFIN expired, we 7247 * will send what we can and go to transmit state. 7248 */ 7249 if (tp->t_flags & TF_FORCEDATA) { 7250 if (sendwin == 0) { 7251 /* 7252 * If we still have some data to send, then clear 7253 * the FIN bit. Usually this would happen below 7254 * when it realizes that we aren't sending all the 7255 * data. However, if we have exactly 1 byte of 7256 * unsent data, then it won't clear the FIN bit 7257 * below, and if we are in persist state, we wind up 7258 * sending the packet without recording that we sent 7259 * the FIN bit. 7260 * 7261 * We can't just blindly clear the FIN bit, because 7262 * if we don't have any more data to send then the 7263 * probe will be the FIN itself. 7264 */ 7265 if (sb_offset < sbused(sb)) 7266 flags &= ~TH_FIN; 7267 sendwin = 1; 7268 } else { 7269 if (rack->rc_in_persist) 7270 rack_exit_persist(tp, rack); 7271 /* 7272 * If we are dropping persist mode then we need to 7273 * correct snd_nxt/snd_max and off. 7274 */ 7275 tp->snd_nxt = tp->snd_max; 7276 sb_offset = tp->snd_nxt - tp->snd_una; 7277 } 7278 } 7279 /* 7280 * If snd_nxt == snd_max and we have transmitted a FIN, the 7281 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 7282 * negative length. This can also occur when TCP opens up its 7283 * congestion window while receiving additional duplicate acks after 7284 * fast-retransmit because TCP will reset snd_nxt to snd_max after 7285 * the fast-retransmit. 7286 * 7287 * In the normal retransmit-FIN-only case, however, snd_nxt will be 7288 * set to snd_una, the sb_offset will be 0, and the length may wind 7289 * up 0. 7290 * 7291 * If sack_rxmit is true we are retransmitting from the scoreboard 7292 * in which case len is already set. 7293 */ 7294 if (sack_rxmit == 0) { 7295 uint32_t avail; 7296 7297 avail = sbavail(sb); 7298 if (SEQ_GT(tp->snd_nxt, tp->snd_una)) 7299 sb_offset = tp->snd_nxt - tp->snd_una; 7300 else 7301 sb_offset = 0; 7302 if (IN_RECOVERY(tp->t_flags) == 0) { 7303 if (rack->r_ctl.rc_tlp_new_data) { 7304 /* TLP is forcing out new data */ 7305 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { 7306 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); 7307 } 7308 if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) 7309 len = tp->snd_wnd; 7310 else 7311 len = rack->r_ctl.rc_tlp_new_data; 7312 rack->r_ctl.rc_tlp_new_data = 0; 7313 doing_tlp = 1; 7314 } else { 7315 if (sendwin > avail) { 7316 /* use the available */ 7317 if (avail > sb_offset) { 7318 len = (int32_t)(avail - sb_offset); 7319 } else { 7320 len = 0; 7321 } 7322 } else { 7323 if (sendwin > sb_offset) { 7324 len = (int32_t)(sendwin - sb_offset); 7325 } else { 7326 len = 0; 7327 } 7328 } 7329 } 7330 } else { 7331 uint32_t outstanding; 7332 7333 /* 7334 * We are inside of a SACK recovery episode and are 7335 * sending new data, having retransmitted all the 7336 * data possible so far in the scoreboard. 7337 */ 7338 outstanding = tp->snd_max - tp->snd_una; 7339 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) { 7340 if (tp->snd_wnd > outstanding) { 7341 len = tp->snd_wnd - outstanding; 7342 /* Check to see if we have the data */ 7343 if (((sb_offset + len) > avail) && 7344 (avail > sb_offset)) 7345 len = avail - sb_offset; 7346 else 7347 len = 0; 7348 } else 7349 len = 0; 7350 } else if (avail > sb_offset) 7351 len = avail - sb_offset; 7352 else 7353 len = 0; 7354 if (len > 0) { 7355 if (len > rack->r_ctl.rc_prr_sndcnt) 7356 len = rack->r_ctl.rc_prr_sndcnt; 7357 7358 if (len > 0) { 7359 sub_from_prr = 1; 7360 counter_u64_add(rack_rtm_prr_newdata, 1); 7361 } 7362 } 7363 if (len > tp->t_maxseg) { 7364 /* 7365 * We should never send more than a MSS when 7366 * retransmitting or sending new data in prr 7367 * mode unless the override flag is on. Most 7368 * likely the PRR algorithm is not going to 7369 * let us send a lot as well :-) 7370 */ 7371 if (rack->r_ctl.rc_prr_sendalot == 0) 7372 len = tp->t_maxseg; 7373 } else if (len < tp->t_maxseg) { 7374 /* 7375 * Do we send any? The idea here is if the 7376 * send empty's the socket buffer we want to 7377 * do it. However if not then lets just wait 7378 * for our prr_sndcnt to get bigger. 7379 */ 7380 long leftinsb; 7381 7382 leftinsb = sbavail(sb) - sb_offset; 7383 if (leftinsb > len) { 7384 /* This send does not empty the sb */ 7385 len = 0; 7386 } 7387 } 7388 } 7389 } 7390 if (prefetch_so_done == 0) { 7391 kern_prefetch(so, &prefetch_so_done); 7392 prefetch_so_done = 1; 7393 } 7394 /* 7395 * Lop off SYN bit if it has already been sent. However, if this is 7396 * SYN-SENT state and if segment contains data and if we don't know 7397 * that foreign host supports TAO, suppress sending segment. 7398 */ 7399 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { 7400 if ((tp->t_state != TCPS_SYN_RECEIVED) && 7401 (tp->t_state != TCPS_SYN_SENT)) 7402 flags &= ~TH_SYN; 7403 #ifdef TCP_RFC7413 7404 /* 7405 * When sending additional segments following a TFO SYN|ACK, 7406 * do not include the SYN bit. 7407 */ 7408 if ((tp->t_flags & TF_FASTOPEN) && 7409 (tp->t_state == TCPS_SYN_RECEIVED)) 7410 flags &= ~TH_SYN; 7411 #endif 7412 sb_offset--, len++; 7413 if (sbavail(sb) == 0) 7414 len = 0; 7415 } 7416 /* 7417 * Be careful not to send data and/or FIN on SYN segments. This 7418 * measure is needed to prevent interoperability problems with not 7419 * fully conformant TCP implementations. 7420 */ 7421 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 7422 len = 0; 7423 flags &= ~TH_FIN; 7424 } 7425 #ifdef TCP_RFC7413 7426 /* 7427 * When retransmitting SYN|ACK on a passively-created TFO socket, 7428 * don't include data, as the presence of data may have caused the 7429 * original SYN|ACK to have been dropped by a middlebox. 7430 */ 7431 if ((tp->t_flags & TF_FASTOPEN) && 7432 ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0))) 7433 len = 0; 7434 #endif 7435 if (len <= 0) { 7436 /* 7437 * If FIN has been sent but not acked, but we haven't been 7438 * called to retransmit, len will be < 0. Otherwise, window 7439 * shrank after we sent into it. If window shrank to 0, 7440 * cancel pending retransmit, pull snd_nxt back to (closed) 7441 * window, and set the persist timer if it isn't already 7442 * going. If the window didn't close completely, just wait 7443 * for an ACK. 7444 * 7445 * We also do a general check here to ensure that we will 7446 * set the persist timer when we have data to send, but a 7447 * 0-byte window. This makes sure the persist timer is set 7448 * even if the packet hits one of the "goto send" lines 7449 * below. 7450 */ 7451 len = 0; 7452 if ((tp->snd_wnd == 0) && 7453 (TCPS_HAVEESTABLISHED(tp->t_state)) && 7454 (sb_offset < (int)sbavail(sb))) { 7455 tp->snd_nxt = tp->snd_una; 7456 rack_enter_persist(tp, rack, cts); 7457 } 7458 } 7459 /* len will be >= 0 after this point. */ 7460 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7461 tcp_sndbuf_autoscale(tp, so, sendwin); 7462 /* 7463 * Decide if we can use TCP Segmentation Offloading (if supported by 7464 * hardware). 7465 * 7466 * TSO may only be used if we are in a pure bulk sending state. The 7467 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 7468 * options prevent using TSO. With TSO the TCP header is the same 7469 * (except for the sequence number) for all generated packets. This 7470 * makes it impossible to transmit any options which vary per 7471 * generated segment or packet. 7472 * 7473 * IPv4 handling has a clear separation of ip options and ip header 7474 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does 7475 * the right thing below to provide length of just ip options and thus 7476 * checking for ipoptlen is enough to decide if ip options are present. 7477 */ 7478 7479 #ifdef INET6 7480 if (isipv6) 7481 ipoptlen = ip6_optlen(tp->t_inpcb); 7482 else 7483 #endif 7484 if (tp->t_inpcb->inp_options) 7485 ipoptlen = tp->t_inpcb->inp_options->m_len - 7486 offsetof(struct ipoption, ipopt_list); 7487 else 7488 ipoptlen = 0; 7489 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7490 /* 7491 * Pre-calculate here as we save another lookup into the darknesses 7492 * of IPsec that way and can actually decide if TSO is ok. 7493 */ 7494 #ifdef INET6 7495 if (isipv6 && IPSEC_ENABLED(ipv6)) 7496 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); 7497 #ifdef INET 7498 else 7499 #endif 7500 #endif /* INET6 */ 7501 #ifdef INET 7502 if (IPSEC_ENABLED(ipv4)) 7503 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); 7504 #endif /* INET */ 7505 #endif 7506 7507 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7508 ipoptlen += ipsec_optlen; 7509 #endif 7510 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && 7511 #ifdef NETFLIX_TCP_O_UDP 7512 (tp->t_port == 0) && 7513 #endif 7514 ((tp->t_flags & TF_SIGNATURE) == 0) && 7515 tp->rcv_numsacks == 0 && sack_rxmit == 0 && 7516 ipoptlen == 0) 7517 tso = 1; 7518 { 7519 uint32_t outstanding; 7520 7521 outstanding = tp->snd_max - tp->snd_una; 7522 if (tp->t_flags & TF_SENTFIN) { 7523 /* 7524 * If we sent a fin, snd_max is 1 higher than 7525 * snd_una 7526 */ 7527 outstanding--; 7528 } 7529 if (outstanding > 0) { 7530 /* 7531 * This is sub-optimal. We only send a stand alone 7532 * FIN on its own segment. 7533 */ 7534 if (flags & TH_FIN) { 7535 flags &= ~TH_FIN; 7536 would_have_fin = 1; 7537 } 7538 } else if (sack_rxmit) { 7539 if ((rsm->r_flags & RACK_HAS_FIN) == 0) 7540 flags &= ~TH_FIN; 7541 } else { 7542 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + 7543 sbused(sb))) 7544 flags &= ~TH_FIN; 7545 } 7546 } 7547 recwin = sbspace(&so->so_rcv); 7548 7549 /* 7550 * Sender silly window avoidance. We transmit under the following 7551 * conditions when len is non-zero: 7552 * 7553 * - We have a full segment (or more with TSO) - This is the last 7554 * buffer in a write()/send() and we are either idle or running 7555 * NODELAY - we've timed out (e.g. persist timer) - we have more 7556 * then 1/2 the maximum send window's worth of data (receiver may be 7557 * limited the window size) - we need to retransmit 7558 */ 7559 if (len) { 7560 if (len >= tp->t_maxseg) { 7561 pass = 1; 7562 goto send; 7563 } 7564 /* 7565 * NOTE! on localhost connections an 'ack' from the remote 7566 * end may occur synchronously with the output and cause us 7567 * to flush a buffer queued with moretocome. XXX 7568 * 7569 */ 7570 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ 7571 (idle || (tp->t_flags & TF_NODELAY)) && 7572 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && 7573 (tp->t_flags & TF_NOPUSH) == 0) { 7574 pass = 2; 7575 goto send; 7576 } 7577 if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 7578 pass = 3; 7579 goto send; 7580 } 7581 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 7582 goto send; 7583 } 7584 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 7585 pass = 4; 7586 goto send; 7587 } 7588 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ 7589 pass = 5; 7590 goto send; 7591 } 7592 if (sack_rxmit) { 7593 pass = 6; 7594 goto send; 7595 } 7596 } 7597 /* 7598 * Sending of standalone window updates. 7599 * 7600 * Window updates are important when we close our window due to a 7601 * full socket buffer and are opening it again after the application 7602 * reads data from it. Once the window has opened again and the 7603 * remote end starts to send again the ACK clock takes over and 7604 * provides the most current window information. 7605 * 7606 * We must avoid the silly window syndrome whereas every read from 7607 * the receive buffer, no matter how small, causes a window update 7608 * to be sent. We also should avoid sending a flurry of window 7609 * updates when the socket buffer had queued a lot of data and the 7610 * application is doing small reads. 7611 * 7612 * Prevent a flurry of pointless window updates by only sending an 7613 * update when we can increase the advertized window by more than 7614 * 1/4th of the socket buffer capacity. When the buffer is getting 7615 * full or is very small be more aggressive and send an update 7616 * whenever we can increase by two mss sized segments. In all other 7617 * situations the ACK's to new incoming data will carry further 7618 * window increases. 7619 * 7620 * Don't send an independent window update if a delayed ACK is 7621 * pending (it will get piggy-backed on it) or the remote side 7622 * already has done a half-close and won't send more data. Skip 7623 * this if the connection is in T/TCP half-open state. 7624 */ 7625 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 7626 !(tp->t_flags & TF_DELACK) && 7627 !TCPS_HAVERCVDFIN(tp->t_state)) { 7628 /* 7629 * "adv" is the amount we could increase the window, taking 7630 * into account that we are limited by TCP_MAXWIN << 7631 * tp->rcv_scale. 7632 */ 7633 int32_t adv; 7634 int oldwin; 7635 7636 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); 7637 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 7638 oldwin = (tp->rcv_adv - tp->rcv_nxt); 7639 adv -= oldwin; 7640 } else 7641 oldwin = 0; 7642 7643 /* 7644 * If the new window size ends up being the same as the old 7645 * size when it is scaled, then don't force a window update. 7646 */ 7647 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) 7648 goto dontupdate; 7649 7650 if (adv >= (int32_t)(2 * tp->t_maxseg) && 7651 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || 7652 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || 7653 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { 7654 pass = 7; 7655 goto send; 7656 } 7657 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 7658 goto send; 7659 } 7660 dontupdate: 7661 7662 /* 7663 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW 7664 * is also a catch-all for the retransmit timer timeout case. 7665 */ 7666 if (tp->t_flags & TF_ACKNOW) { 7667 pass = 8; 7668 goto send; 7669 } 7670 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { 7671 pass = 9; 7672 goto send; 7673 } 7674 if (SEQ_GT(tp->snd_up, tp->snd_una)) { 7675 pass = 10; 7676 goto send; 7677 } 7678 /* 7679 * If our state indicates that FIN should be sent and we have not 7680 * yet done so, then we need to send. 7681 */ 7682 if (flags & TH_FIN) { 7683 if ((tp->t_flags & TF_SENTFIN) || 7684 (((tp->t_flags & TF_SENTFIN) == 0) && 7685 (tp->snd_nxt == tp->snd_una))) { 7686 pass = 11; 7687 goto send; 7688 } 7689 } 7690 /* 7691 * No reason to send a segment, just return. 7692 */ 7693 just_return: 7694 SOCKBUF_UNLOCK(sb); 7695 just_return_nolock: 7696 if (tot_len_this_send == 0) 7697 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); 7698 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 7699 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); 7700 tp->t_flags &= ~TF_FORCEDATA; 7701 return (0); 7702 7703 send: 7704 if (doing_tlp == 0) { 7705 /* 7706 * Data not a TLP, and its not the rxt firing. If it is the 7707 * rxt firing, we want to leave the tlp_in_progress flag on 7708 * so we don't send another TLP. It has to be a rack timer 7709 * or normal send (response to acked data) to clear the tlp 7710 * in progress flag. 7711 */ 7712 rack->rc_tlp_in_progress = 0; 7713 } 7714 SOCKBUF_LOCK_ASSERT(sb); 7715 if (len > 0) { 7716 if (len >= tp->t_maxseg) 7717 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 7718 else 7719 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 7720 } 7721 /* 7722 * Before ESTABLISHED, force sending of initial options unless TCP 7723 * set not to do any options. NOTE: we assume that the IP/TCP header 7724 * plus TCP options always fit in a single mbuf, leaving room for a 7725 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 7726 * + optlen <= MCLBYTES 7727 */ 7728 optlen = 0; 7729 #ifdef INET6 7730 if (isipv6) 7731 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 7732 else 7733 #endif 7734 hdrlen = sizeof(struct tcpiphdr); 7735 7736 /* 7737 * Compute options for segment. We only have to care about SYN and 7738 * established connection segments. Options for SYN-ACK segments 7739 * are handled in TCP syncache. 7740 */ 7741 to.to_flags = 0; 7742 if ((tp->t_flags & TF_NOOPT) == 0) { 7743 /* Maximum segment size. */ 7744 if (flags & TH_SYN) { 7745 tp->snd_nxt = tp->iss; 7746 to.to_mss = tcp_mssopt(&inp->inp_inc); 7747 #ifdef NETFLIX_TCP_O_UDP 7748 if (tp->t_port) 7749 to.to_mss -= V_tcp_udp_tunneling_overhead; 7750 #endif 7751 to.to_flags |= TOF_MSS; 7752 #ifdef TCP_RFC7413 7753 /* 7754 * Only include the TFO option on the first 7755 * transmission of the SYN|ACK on a 7756 * passively-created TFO socket, as the presence of 7757 * the TFO option may have caused the original 7758 * SYN|ACK to have been dropped by a middlebox. 7759 */ 7760 if ((tp->t_flags & TF_FASTOPEN) && 7761 (tp->t_state == TCPS_SYN_RECEIVED) && 7762 (tp->t_rxtshift == 0)) { 7763 to.to_tfo_len = TCP_FASTOPEN_MAX_COOKIE_LEN; 7764 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie; 7765 to.to_flags |= TOF_FASTOPEN; 7766 } 7767 #endif 7768 } 7769 /* Window scaling. */ 7770 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 7771 to.to_wscale = tp->request_r_scale; 7772 to.to_flags |= TOF_SCALE; 7773 } 7774 /* Timestamps. */ 7775 if ((tp->t_flags & TF_RCVD_TSTMP) || 7776 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 7777 to.to_tsval = cts + tp->ts_offset; 7778 to.to_tsecr = tp->ts_recent; 7779 to.to_flags |= TOF_TS; 7780 } 7781 /* Set receive buffer autosizing timestamp. */ 7782 if (tp->rfbuf_ts == 0 && 7783 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 7784 tp->rfbuf_ts = tcp_ts_getticks(); 7785 /* Selective ACK's. */ 7786 if (flags & TH_SYN) 7787 to.to_flags |= TOF_SACKPERM; 7788 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 7789 tp->rcv_numsacks > 0) { 7790 to.to_flags |= TOF_SACK; 7791 to.to_nsacks = tp->rcv_numsacks; 7792 to.to_sacks = (u_char *)tp->sackblks; 7793 } 7794 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 7795 /* TCP-MD5 (RFC2385). */ 7796 if (tp->t_flags & TF_SIGNATURE) 7797 to.to_flags |= TOF_SIGNATURE; 7798 #endif /* TCP_SIGNATURE */ 7799 7800 /* Processing the options. */ 7801 hdrlen += optlen = tcp_addoptions(&to, opt); 7802 } 7803 #ifdef NETFLIX_TCP_O_UDP 7804 if (tp->t_port) { 7805 if (V_tcp_udp_tunneling_port == 0) { 7806 /* The port was removed?? */ 7807 SOCKBUF_UNLOCK(&so->so_snd); 7808 return (EHOSTUNREACH); 7809 } 7810 hdrlen += sizeof(struct udphdr); 7811 } 7812 #endif 7813 ipoptlen = 0; 7814 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 7815 ipoptlen += ipsec_optlen; 7816 #endif 7817 7818 /* 7819 * Adjust data length if insertion of options will bump the packet 7820 * length beyond the t_maxseg length. Clear the FIN bit because we 7821 * cut off the tail of the segment. 7822 */ 7823 if (len + optlen + ipoptlen > tp->t_maxseg) { 7824 if (flags & TH_FIN) { 7825 would_have_fin = 1; 7826 flags &= ~TH_FIN; 7827 } 7828 if (tso) { 7829 uint32_t if_hw_tsomax; 7830 uint32_t moff; 7831 int32_t max_len; 7832 7833 /* extract TSO information */ 7834 if_hw_tsomax = tp->t_tsomax; 7835 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 7836 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 7837 KASSERT(ipoptlen == 0, 7838 ("%s: TSO can't do IP options", __func__)); 7839 7840 /* 7841 * Check if we should limit by maximum payload 7842 * length: 7843 */ 7844 if (if_hw_tsomax != 0) { 7845 /* compute maximum TSO length */ 7846 max_len = (if_hw_tsomax - hdrlen - 7847 max_linkhdr); 7848 if (max_len <= 0) { 7849 len = 0; 7850 } else if (len > max_len) { 7851 sendalot = 1; 7852 len = max_len; 7853 } 7854 } 7855 /* 7856 * Prevent the last segment from being fractional 7857 * unless the send sockbuf can be emptied: 7858 */ 7859 max_len = (tp->t_maxseg - optlen); 7860 if ((sb_offset + len) < sbavail(sb)) { 7861 moff = len % (u_int)max_len; 7862 if (moff != 0) { 7863 len -= moff; 7864 sendalot = 1; 7865 } 7866 } 7867 /* 7868 * In case there are too many small fragments don't 7869 * use TSO: 7870 */ 7871 if (len <= max_len) { 7872 len = max_len; 7873 sendalot = 1; 7874 tso = 0; 7875 } 7876 /* 7877 * Send the FIN in a separate segment after the bulk 7878 * sending is done. We don't trust the TSO 7879 * implementations to clear the FIN flag on all but 7880 * the last segment. 7881 */ 7882 if (tp->t_flags & TF_NEEDFIN) 7883 sendalot = 1; 7884 7885 } else { 7886 len = tp->t_maxseg - optlen - ipoptlen; 7887 sendalot = 1; 7888 } 7889 } else 7890 tso = 0; 7891 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 7892 ("%s: len > IP_MAXPACKET", __func__)); 7893 #ifdef DIAGNOSTIC 7894 #ifdef INET6 7895 if (max_linkhdr + hdrlen > MCLBYTES) 7896 #else 7897 if (max_linkhdr + hdrlen > MHLEN) 7898 #endif 7899 panic("tcphdr too big"); 7900 #endif 7901 7902 /* 7903 * This KASSERT is here to catch edge cases at a well defined place. 7904 * Before, those had triggered (random) panic conditions further 7905 * down. 7906 */ 7907 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 7908 if ((len == 0) && 7909 (flags & TH_FIN) && 7910 (sbused(sb))) { 7911 /* 7912 * We have outstanding data, don't send a fin by itself!. 7913 */ 7914 goto just_return; 7915 } 7916 /* 7917 * Grab a header mbuf, attaching a copy of data to be transmitted, 7918 * and initialize the header from the template for sends on this 7919 * connection. 7920 */ 7921 if (len) { 7922 uint32_t max_val; 7923 uint32_t moff; 7924 7925 if (rack->rc_pace_max_segs) 7926 max_val = rack->rc_pace_max_segs * tp->t_maxseg; 7927 else 7928 max_val = len; 7929 /* 7930 * We allow a limit on sending with hptsi. 7931 */ 7932 if (len > max_val) { 7933 len = max_val; 7934 } 7935 #ifdef INET6 7936 if (MHLEN < hdrlen + max_linkhdr) 7937 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 7938 else 7939 #endif 7940 m = m_gethdr(M_NOWAIT, MT_DATA); 7941 7942 if (m == NULL) { 7943 SOCKBUF_UNLOCK(sb); 7944 error = ENOBUFS; 7945 sack_rxmit = 0; 7946 goto out; 7947 } 7948 m->m_data += max_linkhdr; 7949 m->m_len = hdrlen; 7950 7951 /* 7952 * Start the m_copy functions from the closest mbuf to the 7953 * sb_offset in the socket buffer chain. 7954 */ 7955 mb = sbsndptr_noadv(sb, sb_offset, &moff); 7956 if (len <= MHLEN - hdrlen - max_linkhdr) { 7957 m_copydata(mb, moff, (int)len, 7958 mtod(m, caddr_t)+hdrlen); 7959 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7960 sbsndptr_adv(sb, mb, len); 7961 m->m_len += len; 7962 } else { 7963 struct sockbuf *msb; 7964 7965 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7966 msb = NULL; 7967 else 7968 msb = sb; 7969 m->m_next = tcp_m_copym(/*tp, */ mb, moff, &len, 7970 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb /*, 0, NULL*/); 7971 if (len <= (tp->t_maxseg - optlen)) { 7972 /* 7973 * Must have ran out of mbufs for the copy 7974 * shorten it to no longer need tso. Lets 7975 * not put on sendalot since we are low on 7976 * mbufs. 7977 */ 7978 tso = 0; 7979 } 7980 if (m->m_next == NULL) { 7981 SOCKBUF_UNLOCK(sb); 7982 (void)m_free(m); 7983 error = ENOBUFS; 7984 sack_rxmit = 0; 7985 goto out; 7986 } 7987 } 7988 if ((tp->t_flags & TF_FORCEDATA) && len == 1) { 7989 TCPSTAT_INC(tcps_sndprobe); 7990 #ifdef NETFLIX_STATS 7991 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) 7992 stats_voi_update_abs_u32(tp->t_stats, 7993 VOI_TCP_RETXPB, len); 7994 else 7995 stats_voi_update_abs_u64(tp->t_stats, 7996 VOI_TCP_TXPB, len); 7997 #endif 7998 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { 7999 if (rsm && (rsm->r_flags & RACK_TLP)) { 8000 /* 8001 * TLP should not count in retran count, but 8002 * in its own bin 8003 */ 8004 /* tp->t_sndtlppack++;*/ 8005 /* tp->t_sndtlpbyte += len;*/ 8006 counter_u64_add(rack_tlp_retran, 1); 8007 counter_u64_add(rack_tlp_retran_bytes, len); 8008 } else { 8009 tp->t_sndrexmitpack++; 8010 TCPSTAT_INC(tcps_sndrexmitpack); 8011 TCPSTAT_ADD(tcps_sndrexmitbyte, len); 8012 } 8013 #ifdef NETFLIX_STATS 8014 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 8015 len); 8016 #endif 8017 } else { 8018 TCPSTAT_INC(tcps_sndpack); 8019 TCPSTAT_ADD(tcps_sndbyte, len); 8020 #ifdef NETFLIX_STATS 8021 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 8022 len); 8023 #endif 8024 } 8025 /* 8026 * If we're sending everything we've got, set PUSH. (This 8027 * will keep happy those implementations which only give 8028 * data to the user when a buffer fills or a PUSH comes in.) 8029 */ 8030 if (sb_offset + len == sbused(sb) && 8031 sbused(sb) && 8032 !(flags & TH_SYN)) 8033 flags |= TH_PUSH; 8034 8035 /* 8036 * Are we doing hptsi, if so we must calculate the slot. We 8037 * only do hptsi in ESTABLISHED and with no RESET being 8038 * sent where we have data to send. 8039 */ 8040 if (((tp->t_state == TCPS_ESTABLISHED) || 8041 (tp->t_state == TCPS_CLOSE_WAIT) || 8042 ((tp->t_state == TCPS_FIN_WAIT_1) && 8043 ((tp->t_flags & TF_SENTFIN) == 0) && 8044 ((flags & TH_FIN) == 0))) && 8045 ((flags & TH_RST) == 0) && 8046 (rack->rc_always_pace)) { 8047 /* 8048 * We use the most optimistic possible cwnd/srtt for 8049 * sending calculations. This will make our 8050 * calculation anticipate getting more through 8051 * quicker then possible. But thats ok we don't want 8052 * the peer to have a gap in data sending. 8053 */ 8054 uint32_t srtt, cwnd, tr_perms = 0; 8055 8056 if (rack->r_ctl.rc_rack_min_rtt) 8057 srtt = rack->r_ctl.rc_rack_min_rtt; 8058 else 8059 srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); 8060 if (rack->r_ctl.rc_rack_largest_cwnd) 8061 cwnd = rack->r_ctl.rc_rack_largest_cwnd; 8062 else 8063 cwnd = tp->snd_cwnd; 8064 tr_perms = cwnd / srtt; 8065 if (tr_perms == 0) { 8066 tr_perms = tp->t_maxseg; 8067 } 8068 tot_len_this_send += len; 8069 /* 8070 * Calculate how long this will take to drain, if 8071 * the calculation comes out to zero, thats ok we 8072 * will use send_a_lot to possibly spin around for 8073 * more increasing tot_len_this_send to the point 8074 * that its going to require a pace, or we hit the 8075 * cwnd. Which in that case we are just waiting for 8076 * a ACK. 8077 */ 8078 slot = tot_len_this_send / tr_perms; 8079 /* Now do we reduce the time so we don't run dry? */ 8080 if (slot && rack->rc_pace_reduce) { 8081 int32_t reduce; 8082 8083 reduce = (slot / rack->rc_pace_reduce); 8084 if (reduce < slot) { 8085 slot -= reduce; 8086 } else 8087 slot = 0; 8088 } 8089 if (rack->r_enforce_min_pace && 8090 (slot == 0) && 8091 (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { 8092 /* We are enforcing a minimum pace time of 1ms */ 8093 slot = rack->r_enforce_min_pace; 8094 } 8095 } 8096 SOCKBUF_UNLOCK(sb); 8097 } else { 8098 SOCKBUF_UNLOCK(sb); 8099 if (tp->t_flags & TF_ACKNOW) 8100 TCPSTAT_INC(tcps_sndacks); 8101 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 8102 TCPSTAT_INC(tcps_sndctrl); 8103 else if (SEQ_GT(tp->snd_up, tp->snd_una)) 8104 TCPSTAT_INC(tcps_sndurg); 8105 else 8106 TCPSTAT_INC(tcps_sndwinup); 8107 8108 m = m_gethdr(M_NOWAIT, MT_DATA); 8109 if (m == NULL) { 8110 error = ENOBUFS; 8111 sack_rxmit = 0; 8112 goto out; 8113 } 8114 #ifdef INET6 8115 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 8116 MHLEN >= hdrlen) { 8117 M_ALIGN(m, hdrlen); 8118 } else 8119 #endif 8120 m->m_data += max_linkhdr; 8121 m->m_len = hdrlen; 8122 } 8123 SOCKBUF_UNLOCK_ASSERT(sb); 8124 m->m_pkthdr.rcvif = (struct ifnet *)0; 8125 #ifdef MAC 8126 mac_inpcb_create_mbuf(inp, m); 8127 #endif 8128 #ifdef INET6 8129 if (isipv6) { 8130 ip6 = mtod(m, struct ip6_hdr *); 8131 #ifdef NETFLIX_TCP_O_UDP 8132 if (tp->t_port) { 8133 udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); 8134 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8135 udp->uh_dport = tp->t_port; 8136 ulen = hdrlen + len - sizeof(struct ip6_hdr); 8137 udp->uh_ulen = htons(ulen); 8138 th = (struct tcphdr *)(udp + 1); 8139 } else 8140 #endif 8141 th = (struct tcphdr *)(ip6 + 1); 8142 tcpip_fillheaders(inp, /*tp->t_port, */ ip6, th); 8143 } else 8144 #endif /* INET6 */ 8145 { 8146 ip = mtod(m, struct ip *); 8147 #ifdef TCPDEBUG 8148 ipov = (struct ipovly *)ip; 8149 #endif 8150 #ifdef NETFLIX_TCP_O_UDP 8151 if (tp->t_port) { 8152 udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); 8153 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 8154 udp->uh_dport = tp->t_port; 8155 ulen = hdrlen + len - sizeof(struct ip); 8156 udp->uh_ulen = htons(ulen); 8157 th = (struct tcphdr *)(udp + 1); 8158 } else 8159 #endif 8160 th = (struct tcphdr *)(ip + 1); 8161 tcpip_fillheaders(inp,/*tp->t_port, */ ip, th); 8162 } 8163 /* 8164 * Fill in fields, remembering maximum advertised window for use in 8165 * delaying messages about window sizes. If resending a FIN, be sure 8166 * not to use a new sequence number. 8167 */ 8168 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && 8169 tp->snd_nxt == tp->snd_max) 8170 tp->snd_nxt--; 8171 /* 8172 * If we are starting a connection, send ECN setup SYN packet. If we 8173 * are on a retransmit, we may resend those bits a number of times 8174 * as per RFC 3168. 8175 */ 8176 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { 8177 if (tp->t_rxtshift >= 1) { 8178 if (tp->t_rxtshift <= V_tcp_ecn_maxretries) 8179 flags |= TH_ECE | TH_CWR; 8180 } else 8181 flags |= TH_ECE | TH_CWR; 8182 } 8183 if (tp->t_state == TCPS_ESTABLISHED && 8184 (tp->t_flags & TF_ECN_PERMIT)) { 8185 /* 8186 * If the peer has ECN, mark data packets with ECN capable 8187 * transmission (ECT). Ignore pure ack packets, 8188 * retransmissions and window probes. 8189 */ 8190 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && 8191 !((tp->t_flags & TF_FORCEDATA) && len == 1)) { 8192 #ifdef INET6 8193 if (isipv6) 8194 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); 8195 else 8196 #endif 8197 ip->ip_tos |= IPTOS_ECN_ECT0; 8198 TCPSTAT_INC(tcps_ecn_ect0); 8199 } 8200 /* 8201 * Reply with proper ECN notifications. 8202 */ 8203 if (tp->t_flags & TF_ECN_SND_CWR) { 8204 flags |= TH_CWR; 8205 tp->t_flags &= ~TF_ECN_SND_CWR; 8206 } 8207 if (tp->t_flags & TF_ECN_SND_ECE) 8208 flags |= TH_ECE; 8209 } 8210 /* 8211 * If we are doing retransmissions, then snd_nxt will not reflect 8212 * the first unsent octet. For ACK only packets, we do not want the 8213 * sequence number of the retransmitted packet, we want the sequence 8214 * number of the next unsent octet. So, if there is no data (and no 8215 * SYN or FIN), use snd_max instead of snd_nxt when filling in 8216 * ti_seq. But if we are in persist state, snd_max might reflect 8217 * one byte beyond the right edge of the window, so use snd_nxt in 8218 * that case, since we know we aren't doing a retransmission. 8219 * (retransmit and persist are mutually exclusive...) 8220 */ 8221 if (sack_rxmit == 0) { 8222 if (len || (flags & (TH_SYN | TH_FIN)) || 8223 rack->rc_in_persist) { 8224 th->th_seq = htonl(tp->snd_nxt); 8225 rack_seq = tp->snd_nxt; 8226 } else if (flags & TH_RST) { 8227 /* 8228 * For a Reset send the last cum ack in sequence 8229 * (this like any other choice may still generate a 8230 * challenge ack, if a ack-update packet is in 8231 * flight). 8232 */ 8233 th->th_seq = htonl(tp->snd_una); 8234 rack_seq = tp->snd_una; 8235 } else { 8236 th->th_seq = htonl(tp->snd_max); 8237 rack_seq = tp->snd_max; 8238 } 8239 } else { 8240 th->th_seq = htonl(rsm->r_start); 8241 rack_seq = rsm->r_start; 8242 } 8243 th->th_ack = htonl(tp->rcv_nxt); 8244 if (optlen) { 8245 bcopy(opt, th + 1, optlen); 8246 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 8247 } 8248 th->th_flags = flags; 8249 /* 8250 * Calculate receive window. Don't shrink window, but avoid silly 8251 * window syndrome. 8252 */ 8253 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && 8254 recwin < (long)tp->t_maxseg) 8255 recwin = 0; 8256 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 8257 recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) 8258 recwin = (long)(tp->rcv_adv - tp->rcv_nxt); 8259 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) 8260 recwin = (long)TCP_MAXWIN << tp->rcv_scale; 8261 8262 /* 8263 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 8264 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 8265 * handled in syncache. 8266 */ 8267 if (flags & TH_SYN) 8268 th->th_win = htons((u_short) 8269 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 8270 else 8271 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 8272 /* 8273 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 8274 * window. This may cause the remote transmitter to stall. This 8275 * flag tells soreceive() to disable delayed acknowledgements when 8276 * draining the buffer. This can occur if the receiver is 8277 * attempting to read more data than can be buffered prior to 8278 * transmitting on the connection. 8279 */ 8280 if (th->th_win == 0) { 8281 tp->t_sndzerowin++; 8282 tp->t_flags |= TF_RXWIN0SENT; 8283 } else 8284 tp->t_flags &= ~TF_RXWIN0SENT; 8285 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { 8286 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); 8287 th->th_flags |= TH_URG; 8288 } else 8289 /* 8290 * If no urgent pointer to send, then we pull the urgent 8291 * pointer to the left edge of the send window so that it 8292 * doesn't drift into the send window on sequence number 8293 * wraparound. 8294 */ 8295 tp->snd_up = tp->snd_una; /* drag it along */ 8296 8297 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 8298 if (to.to_flags & TOF_SIGNATURE) { 8299 /* 8300 * Calculate MD5 signature and put it into the place 8301 * determined before. 8302 * NOTE: since TCP options buffer doesn't point into 8303 * mbuf's data, calculate offset and use it. 8304 */ 8305 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 8306 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 8307 /* 8308 * Do not send segment if the calculation of MD5 8309 * digest has failed. 8310 */ 8311 goto out; 8312 } 8313 } 8314 #endif 8315 8316 /* 8317 * Put TCP length in extended header, and then checksum extended 8318 * header and data. 8319 */ 8320 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 8321 #ifdef INET6 8322 if (isipv6) { 8323 /* 8324 * ip6_plen is not need to be filled now, and will be filled 8325 * in ip6_output. 8326 */ 8327 #ifdef NETFLIX_TCP_O_UDP 8328 if (tp->t_port) { 8329 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 8330 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8331 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 8332 th->th_sum = htons(0); 8333 UDPSTAT_INC(udps_opackets); 8334 } else { 8335 #endif 8336 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 8337 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8338 th->th_sum = in6_cksum_pseudo(ip6, 8339 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 8340 0); 8341 #ifdef NETFLIX_TCP_O_UDP 8342 } 8343 #endif 8344 } 8345 #endif 8346 #if defined(INET6) && defined(INET) 8347 else 8348 #endif 8349 #ifdef INET 8350 { 8351 #ifdef NETFLIX_TCP_O_UDP 8352 if (tp->t_port) { 8353 m->m_pkthdr.csum_flags = CSUM_UDP; 8354 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 8355 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 8356 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 8357 th->th_sum = htons(0); 8358 UDPSTAT_INC(udps_opackets); 8359 } else { 8360 #endif 8361 m->m_pkthdr.csum_flags = CSUM_TCP; 8362 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 8363 th->th_sum = in_pseudo(ip->ip_src.s_addr, 8364 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 8365 IPPROTO_TCP + len + optlen)); 8366 #ifdef NETFLIX_TCP_O_UDP 8367 } 8368 #endif 8369 /* IP version must be set here for ipv4/ipv6 checking later */ 8370 KASSERT(ip->ip_v == IPVERSION, 8371 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 8372 } 8373 #endif 8374 8375 /* 8376 * Enable TSO and specify the size of the segments. The TCP pseudo 8377 * header checksum is always provided. XXX: Fixme: This is currently 8378 * not the case for IPv6. 8379 */ 8380 if (tso) { 8381 KASSERT(len > tp->t_maxseg - optlen, 8382 ("%s: len <= tso_segsz", __func__)); 8383 m->m_pkthdr.csum_flags |= CSUM_TSO; 8384 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; 8385 } 8386 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 8387 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), 8388 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", 8389 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); 8390 #else 8391 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), 8392 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", 8393 __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); 8394 #endif 8395 8396 #ifdef TCP_HHOOK 8397 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ 8398 hhook_run_tcp_est_out(tp, th, &to, len, tso); 8399 #endif 8400 8401 #ifdef TCPDEBUG 8402 /* 8403 * Trace. 8404 */ 8405 if (so->so_options & SO_DEBUG) { 8406 u_short save = 0; 8407 8408 #ifdef INET6 8409 if (!isipv6) 8410 #endif 8411 { 8412 save = ipov->ih_len; 8413 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 8414 * (th->th_off << 2) */ ); 8415 } 8416 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 8417 #ifdef INET6 8418 if (!isipv6) 8419 #endif 8420 ipov->ih_len = save; 8421 } 8422 #endif /* TCPDEBUG */ 8423 8424 /* We're getting ready to send; log now. */ 8425 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 8426 union tcp_log_stackspecific log; 8427 8428 memset(&log.u_bbr, 0, sizeof(log.u_bbr)); 8429 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; 8430 log.u_bbr.ininput = rack->rc_inp->inp_in_input; 8431 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; 8432 if (rsm || sack_rxmit) { 8433 log.u_bbr.flex8 = 1; 8434 } else { 8435 log.u_bbr.flex8 = 0; 8436 } 8437 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 8438 len, &log, false, NULL, NULL, 0, NULL); 8439 } else 8440 lgb = NULL; 8441 8442 /* 8443 * Fill in IP length and desired time to live and send to IP level. 8444 * There should be a better way to handle ttl and tos; we could keep 8445 * them in the template, but need a way to checksum without them. 8446 */ 8447 /* 8448 * m->m_pkthdr.len should have been set before cksum calcuration, 8449 * because in6_cksum() need it. 8450 */ 8451 #ifdef INET6 8452 if (isipv6) { 8453 /* 8454 * we separately set hoplimit for every segment, since the 8455 * user might want to change the value via setsockopt. Also, 8456 * desired default hop limit might be changed via Neighbor 8457 * Discovery. 8458 */ 8459 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 8460 8461 /* 8462 * Set the packet size here for the benefit of DTrace 8463 * probes. ip6_output() will set it properly; it's supposed 8464 * to include the option header lengths as well. 8465 */ 8466 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 8467 8468 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) 8469 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8470 else 8471 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8472 8473 if (tp->t_state == TCPS_SYN_SENT) 8474 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 8475 8476 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 8477 /* TODO: IPv6 IP6TOS_ECT bit on */ 8478 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, 8479 &inp->inp_route6, 8480 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 8481 NULL, NULL, inp); 8482 8483 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) 8484 mtu = inp->inp_route6.ro_rt->rt_mtu; 8485 } 8486 #endif /* INET6 */ 8487 #if defined(INET) && defined(INET6) 8488 else 8489 #endif 8490 #ifdef INET 8491 { 8492 ip->ip_len = htons(m->m_pkthdr.len); 8493 #ifdef INET6 8494 if (inp->inp_vflag & INP_IPV6PROTO) 8495 ip->ip_ttl = in6_selecthlim(inp, NULL); 8496 #endif /* INET6 */ 8497 /* 8498 * If we do path MTU discovery, then we set DF on every 8499 * packet. This might not be the best thing to do according 8500 * to RFC3390 Section 2. However the tcp hostcache migitates 8501 * the problem so it affects only the first tcp connection 8502 * with a host. 8503 * 8504 * NB: Don't set DF on small MTU/MSS to have a safe 8505 * fallback. 8506 */ 8507 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 8508 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 8509 if (tp->t_port == 0 || len < V_tcp_minmss) { 8510 ip->ip_off |= htons(IP_DF); 8511 } 8512 } else { 8513 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 8514 } 8515 8516 if (tp->t_state == TCPS_SYN_SENT) 8517 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 8518 8519 TCP_PROBE5(send, NULL, tp, ip, tp, th); 8520 8521 error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, 8522 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, 8523 inp); 8524 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) 8525 mtu = inp->inp_route.ro_rt->rt_mtu; 8526 } 8527 #endif /* INET */ 8528 8529 out: 8530 if (lgb) { 8531 lgb->tlb_errno = error; 8532 lgb = NULL; 8533 } 8534 /* 8535 * In transmit state, time the transmission and arrange for the 8536 * retransmit. In persist state, just set snd_max. 8537 */ 8538 if (error == 0) { 8539 if (TCPS_HAVEESTABLISHED(tp->t_state) && 8540 (tp->t_flags & TF_SACK_PERMIT) && 8541 tp->rcv_numsacks > 0) 8542 tcp_clean_dsack_blocks(tp); 8543 if (len == 0) 8544 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); 8545 else if (len == 1) { 8546 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); 8547 } else if (len > 1) { 8548 int idx; 8549 8550 idx = (len / tp->t_maxseg) + 3; 8551 if (idx >= TCP_MSS_ACCT_ATIMER) 8552 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); 8553 else 8554 counter_u64_add(rack_out_size[idx], 1); 8555 } 8556 } 8557 if (sub_from_prr && (error == 0)) { 8558 if (rack->r_ctl.rc_prr_sndcnt >= len) 8559 rack->r_ctl.rc_prr_sndcnt -= len; 8560 else 8561 rack->r_ctl.rc_prr_sndcnt = 0; 8562 } 8563 sub_from_prr = 0; 8564 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, 8565 pass, rsm); 8566 if ((tp->t_flags & TF_FORCEDATA) == 0 || 8567 (rack->rc_in_persist == 0)) { 8568 #ifdef NETFLIX_STATS 8569 tcp_seq startseq = tp->snd_nxt; 8570 #endif 8571 /* 8572 * Advance snd_nxt over sequence space of this segment. 8573 */ 8574 if (error) 8575 /* We don't log or do anything with errors */ 8576 goto timer; 8577 8578 if (flags & (TH_SYN | TH_FIN)) { 8579 if (flags & TH_SYN) 8580 tp->snd_nxt++; 8581 if (flags & TH_FIN) { 8582 tp->snd_nxt++; 8583 tp->t_flags |= TF_SENTFIN; 8584 } 8585 } 8586 /* In the ENOBUFS case we do *not* update snd_max */ 8587 if (sack_rxmit) 8588 goto timer; 8589 8590 tp->snd_nxt += len; 8591 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { 8592 if (tp->snd_una == tp->snd_max) { 8593 /* 8594 * Update the time we just added data since 8595 * none was outstanding. 8596 */ 8597 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8598 tp->t_acktime = ticks; 8599 } 8600 tp->snd_max = tp->snd_nxt; 8601 #ifdef NETFLIX_STATS 8602 if (!(tp->t_flags & TF_GPUTINPROG) && len) { 8603 tp->t_flags |= TF_GPUTINPROG; 8604 tp->gput_seq = startseq; 8605 tp->gput_ack = startseq + 8606 ulmin(sbavail(sb) - sb_offset, sendwin); 8607 tp->gput_ts = tcp_ts_getticks(); 8608 } 8609 #endif 8610 } 8611 /* 8612 * Set retransmit timer if not currently set, and not doing 8613 * a pure ack or a keep-alive probe. Initial value for 8614 * retransmit timer is smoothed round-trip time + 2 * 8615 * round-trip time variance. Initialize shift counter which 8616 * is used for backoff of retransmit time. 8617 */ 8618 timer: 8619 if ((tp->snd_wnd == 0) && 8620 TCPS_HAVEESTABLISHED(tp->t_state)) { 8621 /* 8622 * If the persists timer was set above (right before 8623 * the goto send), and still needs to be on. Lets 8624 * make sure all is canceled. If the persist timer 8625 * is not running, we want to get it up. 8626 */ 8627 if (rack->rc_in_persist == 0) { 8628 rack_enter_persist(tp, rack, cts); 8629 } 8630 } 8631 } else { 8632 /* 8633 * Persist case, update snd_max but since we are in persist 8634 * mode (no window) we do not update snd_nxt. 8635 */ 8636 int32_t xlen = len; 8637 8638 if (error) 8639 goto nomore; 8640 8641 if (flags & TH_SYN) 8642 ++xlen; 8643 if (flags & TH_FIN) { 8644 ++xlen; 8645 tp->t_flags |= TF_SENTFIN; 8646 } 8647 /* In the ENOBUFS case we do *not* update snd_max */ 8648 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { 8649 if (tp->snd_una == tp->snd_max) { 8650 /* 8651 * Update the time we just added data since 8652 * none was outstanding. 8653 */ 8654 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); 8655 tp->t_acktime = ticks; 8656 } 8657 tp->snd_max = tp->snd_nxt + len; 8658 } 8659 } 8660 nomore: 8661 if (error) { 8662 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ 8663 /* 8664 * Failures do not advance the seq counter above. For the 8665 * case of ENOBUFS we will fall out and retry in 1ms with 8666 * the hpts. Everything else will just have to retransmit 8667 * with the timer. 8668 * 8669 * In any case, we do not want to loop around for another 8670 * send without a good reason. 8671 */ 8672 sendalot = 0; 8673 switch (error) { 8674 case EPERM: 8675 tp->t_flags &= ~TF_FORCEDATA; 8676 tp->t_softerror = error; 8677 return (error); 8678 case ENOBUFS: 8679 if (slot == 0) { 8680 /* 8681 * Pace us right away to retry in a some 8682 * time 8683 */ 8684 slot = 1 + rack->rc_enobuf; 8685 if (rack->rc_enobuf < 255) 8686 rack->rc_enobuf++; 8687 if (slot > (rack->rc_rack_rtt / 2)) { 8688 slot = rack->rc_rack_rtt / 2; 8689 } 8690 if (slot < 10) 8691 slot = 10; 8692 } 8693 counter_u64_add(rack_saw_enobuf, 1); 8694 error = 0; 8695 goto enobufs; 8696 case EMSGSIZE: 8697 /* 8698 * For some reason the interface we used initially 8699 * to send segments changed to another or lowered 8700 * its MTU. If TSO was active we either got an 8701 * interface without TSO capabilits or TSO was 8702 * turned off. If we obtained mtu from ip_output() 8703 * then update it and try again. 8704 */ 8705 if (tso) 8706 tp->t_flags &= ~TF_TSO; 8707 if (mtu != 0) { 8708 tcp_mss_update(tp, -1, mtu, NULL, NULL); 8709 goto again; 8710 } 8711 slot = 10; 8712 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8713 tp->t_flags &= ~TF_FORCEDATA; 8714 return (error); 8715 case ENETUNREACH: 8716 counter_u64_add(rack_saw_enetunreach, 1); 8717 case EHOSTDOWN: 8718 case EHOSTUNREACH: 8719 case ENETDOWN: 8720 if (TCPS_HAVERCVDSYN(tp->t_state)) { 8721 tp->t_softerror = error; 8722 } 8723 /* FALLTHROUGH */ 8724 default: 8725 slot = 10; 8726 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); 8727 tp->t_flags &= ~TF_FORCEDATA; 8728 return (error); 8729 } 8730 } else { 8731 rack->rc_enobuf = 0; 8732 } 8733 TCPSTAT_INC(tcps_sndtotal); 8734 8735 /* 8736 * Data sent (as far as we can tell). If this advertises a larger 8737 * window than any other segment, then remember the size of the 8738 * advertised window. Any pending ACK has now been sent. 8739 */ 8740 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 8741 tp->rcv_adv = tp->rcv_nxt + recwin; 8742 tp->last_ack_sent = tp->rcv_nxt; 8743 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 8744 enobufs: 8745 rack->r_tlp_running = 0; 8746 if ((flags & TH_RST) || (would_have_fin == 1)) { 8747 /* 8748 * We don't send again after a RST. We also do *not* send 8749 * again if we would have had a find, but now have 8750 * outstanding data. 8751 */ 8752 slot = 0; 8753 sendalot = 0; 8754 } 8755 if (slot) { 8756 /* set the rack tcb into the slot N */ 8757 counter_u64_add(rack_paced_segments, 1); 8758 } else if (sendalot) { 8759 if (len) 8760 counter_u64_add(rack_unpaced_segments, 1); 8761 sack_rxmit = 0; 8762 tp->t_flags &= ~TF_FORCEDATA; 8763 goto again; 8764 } else if (len) { 8765 counter_u64_add(rack_unpaced_segments, 1); 8766 } 8767 tp->t_flags &= ~TF_FORCEDATA; 8768 rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); 8769 return (error); 8770 } 8771 8772 /* 8773 * rack_ctloutput() must drop the inpcb lock before performing copyin on 8774 * socket option arguments. When it re-acquires the lock after the copy, it 8775 * has to revalidate that the connection is still valid for the socket 8776 * option. 8777 */ 8778 static int 8779 rack_set_sockopt(struct socket *so, struct sockopt *sopt, 8780 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8781 { 8782 int32_t error = 0, optval; 8783 8784 switch (sopt->sopt_name) { 8785 case TCP_RACK_PROP_RATE: 8786 case TCP_RACK_PROP: 8787 case TCP_RACK_TLP_REDUCE: 8788 case TCP_RACK_EARLY_RECOV: 8789 case TCP_RACK_PACE_ALWAYS: 8790 case TCP_DELACK: 8791 case TCP_RACK_PACE_REDUCE: 8792 case TCP_RACK_PACE_MAX_SEG: 8793 case TCP_RACK_PRR_SENDALOT: 8794 case TCP_RACK_MIN_TO: 8795 case TCP_RACK_EARLY_SEG: 8796 case TCP_RACK_REORD_THRESH: 8797 case TCP_RACK_REORD_FADE: 8798 case TCP_RACK_TLP_THRESH: 8799 case TCP_RACK_PKT_DELAY: 8800 case TCP_RACK_TLP_USE: 8801 case TCP_RACK_TLP_INC_VAR: 8802 case TCP_RACK_IDLE_REDUCE_HIGH: 8803 case TCP_RACK_MIN_PACE: 8804 case TCP_RACK_MIN_PACE_SEG: 8805 case TCP_BBR_RACK_RTT_USE: 8806 case TCP_DATA_AFTER_CLOSE: 8807 break; 8808 default: 8809 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8810 break; 8811 } 8812 INP_WUNLOCK(inp); 8813 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 8814 if (error) 8815 return (error); 8816 INP_WLOCK(inp); 8817 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 8818 INP_WUNLOCK(inp); 8819 return (ECONNRESET); 8820 } 8821 tp = intotcpcb(inp); 8822 rack = (struct tcp_rack *)tp->t_fb_ptr; 8823 switch (sopt->sopt_name) { 8824 case TCP_RACK_PROP_RATE: 8825 if ((optval <= 0) || (optval >= 100)) { 8826 error = EINVAL; 8827 break; 8828 } 8829 RACK_OPTS_INC(tcp_rack_prop_rate); 8830 rack->r_ctl.rc_prop_rate = optval; 8831 break; 8832 case TCP_RACK_TLP_USE: 8833 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { 8834 error = EINVAL; 8835 break; 8836 } 8837 RACK_OPTS_INC(tcp_tlp_use); 8838 rack->rack_tlp_threshold_use = optval; 8839 break; 8840 case TCP_RACK_PROP: 8841 /* RACK proportional rate reduction (bool) */ 8842 RACK_OPTS_INC(tcp_rack_prop); 8843 rack->r_ctl.rc_prop_reduce = optval; 8844 break; 8845 case TCP_RACK_TLP_REDUCE: 8846 /* RACK TLP cwnd reduction (bool) */ 8847 RACK_OPTS_INC(tcp_rack_tlp_reduce); 8848 rack->r_ctl.rc_tlp_cwnd_reduce = optval; 8849 break; 8850 case TCP_RACK_EARLY_RECOV: 8851 /* Should recovery happen early (bool) */ 8852 RACK_OPTS_INC(tcp_rack_early_recov); 8853 rack->r_ctl.rc_early_recovery = optval; 8854 break; 8855 case TCP_RACK_PACE_ALWAYS: 8856 /* Use the always pace method (bool) */ 8857 RACK_OPTS_INC(tcp_rack_pace_always); 8858 if (optval > 0) 8859 rack->rc_always_pace = 1; 8860 else 8861 rack->rc_always_pace = 0; 8862 break; 8863 case TCP_RACK_PACE_REDUCE: 8864 /* RACK Hptsi reduction factor (divisor) */ 8865 RACK_OPTS_INC(tcp_rack_pace_reduce); 8866 if (optval) 8867 /* Must be non-zero */ 8868 rack->rc_pace_reduce = optval; 8869 else 8870 error = EINVAL; 8871 break; 8872 case TCP_RACK_PACE_MAX_SEG: 8873 /* Max segments in a pace */ 8874 RACK_OPTS_INC(tcp_rack_max_seg); 8875 rack->rc_pace_max_segs = optval; 8876 break; 8877 case TCP_RACK_PRR_SENDALOT: 8878 /* Allow PRR to send more than one seg */ 8879 RACK_OPTS_INC(tcp_rack_prr_sendalot); 8880 rack->r_ctl.rc_prr_sendalot = optval; 8881 break; 8882 case TCP_RACK_MIN_TO: 8883 /* Minimum time between rack t-o's in ms */ 8884 RACK_OPTS_INC(tcp_rack_min_to); 8885 rack->r_ctl.rc_min_to = optval; 8886 break; 8887 case TCP_RACK_EARLY_SEG: 8888 /* If early recovery max segments */ 8889 RACK_OPTS_INC(tcp_rack_early_seg); 8890 rack->r_ctl.rc_early_recovery_segs = optval; 8891 break; 8892 case TCP_RACK_REORD_THRESH: 8893 /* RACK reorder threshold (shift amount) */ 8894 RACK_OPTS_INC(tcp_rack_reord_thresh); 8895 if ((optval > 0) && (optval < 31)) 8896 rack->r_ctl.rc_reorder_shift = optval; 8897 else 8898 error = EINVAL; 8899 break; 8900 case TCP_RACK_REORD_FADE: 8901 /* Does reordering fade after ms time */ 8902 RACK_OPTS_INC(tcp_rack_reord_fade); 8903 rack->r_ctl.rc_reorder_fade = optval; 8904 break; 8905 case TCP_RACK_TLP_THRESH: 8906 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 8907 RACK_OPTS_INC(tcp_rack_tlp_thresh); 8908 if (optval) 8909 rack->r_ctl.rc_tlp_threshold = optval; 8910 else 8911 error = EINVAL; 8912 break; 8913 case TCP_RACK_PKT_DELAY: 8914 /* RACK added ms i.e. rack-rtt + reord + N */ 8915 RACK_OPTS_INC(tcp_rack_pkt_delay); 8916 rack->r_ctl.rc_pkt_delay = optval; 8917 break; 8918 case TCP_RACK_TLP_INC_VAR: 8919 /* Does TLP include rtt variance in t-o */ 8920 RACK_OPTS_INC(tcp_rack_tlp_inc_var); 8921 rack->r_ctl.rc_prr_inc_var = optval; 8922 break; 8923 case TCP_RACK_IDLE_REDUCE_HIGH: 8924 RACK_OPTS_INC(tcp_rack_idle_reduce_high); 8925 if (optval) 8926 rack->r_idle_reduce_largest = 1; 8927 else 8928 rack->r_idle_reduce_largest = 0; 8929 break; 8930 case TCP_DELACK: 8931 if (optval == 0) 8932 tp->t_delayed_ack = 0; 8933 else 8934 tp->t_delayed_ack = 1; 8935 if (tp->t_flags & TF_DELACK) { 8936 tp->t_flags &= ~TF_DELACK; 8937 tp->t_flags |= TF_ACKNOW; 8938 rack_output(tp); 8939 } 8940 break; 8941 case TCP_RACK_MIN_PACE: 8942 RACK_OPTS_INC(tcp_rack_min_pace); 8943 if (optval > 3) 8944 rack->r_enforce_min_pace = 3; 8945 else 8946 rack->r_enforce_min_pace = optval; 8947 break; 8948 case TCP_RACK_MIN_PACE_SEG: 8949 RACK_OPTS_INC(tcp_rack_min_pace_seg); 8950 if (optval >= 16) 8951 rack->r_min_pace_seg_thresh = 15; 8952 else 8953 rack->r_min_pace_seg_thresh = optval; 8954 break; 8955 case TCP_BBR_RACK_RTT_USE: 8956 if ((optval != USE_RTT_HIGH) && 8957 (optval != USE_RTT_LOW) && 8958 (optval != USE_RTT_AVG)) 8959 error = EINVAL; 8960 else 8961 rack->r_ctl.rc_rate_sample_method = optval; 8962 break; 8963 case TCP_DATA_AFTER_CLOSE: 8964 if (optval) 8965 rack->rc_allow_data_af_clo = 1; 8966 else 8967 rack->rc_allow_data_af_clo = 0; 8968 break; 8969 default: 8970 return (tcp_default_ctloutput(so, sopt, inp, tp)); 8971 break; 8972 } 8973 /* tcp_log_socket_option(tp, sopt->sopt_name, optval, error);*/ 8974 INP_WUNLOCK(inp); 8975 return (error); 8976 } 8977 8978 static int 8979 rack_get_sockopt(struct socket *so, struct sockopt *sopt, 8980 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) 8981 { 8982 int32_t error, optval; 8983 8984 /* 8985 * Because all our options are either boolean or an int, we can just 8986 * pull everything into optval and then unlock and copy. If we ever 8987 * add a option that is not a int, then this will have quite an 8988 * impact to this routine. 8989 */ 8990 switch (sopt->sopt_name) { 8991 case TCP_RACK_PROP_RATE: 8992 optval = rack->r_ctl.rc_prop_rate; 8993 break; 8994 case TCP_RACK_PROP: 8995 /* RACK proportional rate reduction (bool) */ 8996 optval = rack->r_ctl.rc_prop_reduce; 8997 break; 8998 case TCP_RACK_TLP_REDUCE: 8999 /* RACK TLP cwnd reduction (bool) */ 9000 optval = rack->r_ctl.rc_tlp_cwnd_reduce; 9001 break; 9002 case TCP_RACK_EARLY_RECOV: 9003 /* Should recovery happen early (bool) */ 9004 optval = rack->r_ctl.rc_early_recovery; 9005 break; 9006 case TCP_RACK_PACE_REDUCE: 9007 /* RACK Hptsi reduction factor (divisor) */ 9008 optval = rack->rc_pace_reduce; 9009 break; 9010 case TCP_RACK_PACE_MAX_SEG: 9011 /* Max segments in a pace */ 9012 optval = rack->rc_pace_max_segs; 9013 break; 9014 case TCP_RACK_PACE_ALWAYS: 9015 /* Use the always pace method */ 9016 optval = rack->rc_always_pace; 9017 break; 9018 case TCP_RACK_PRR_SENDALOT: 9019 /* Allow PRR to send more than one seg */ 9020 optval = rack->r_ctl.rc_prr_sendalot; 9021 break; 9022 case TCP_RACK_MIN_TO: 9023 /* Minimum time between rack t-o's in ms */ 9024 optval = rack->r_ctl.rc_min_to; 9025 break; 9026 case TCP_RACK_EARLY_SEG: 9027 /* If early recovery max segments */ 9028 optval = rack->r_ctl.rc_early_recovery_segs; 9029 break; 9030 case TCP_RACK_REORD_THRESH: 9031 /* RACK reorder threshold (shift amount) */ 9032 optval = rack->r_ctl.rc_reorder_shift; 9033 break; 9034 case TCP_RACK_REORD_FADE: 9035 /* Does reordering fade after ms time */ 9036 optval = rack->r_ctl.rc_reorder_fade; 9037 break; 9038 case TCP_RACK_TLP_THRESH: 9039 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 9040 optval = rack->r_ctl.rc_tlp_threshold; 9041 break; 9042 case TCP_RACK_PKT_DELAY: 9043 /* RACK added ms i.e. rack-rtt + reord + N */ 9044 optval = rack->r_ctl.rc_pkt_delay; 9045 break; 9046 case TCP_RACK_TLP_USE: 9047 optval = rack->rack_tlp_threshold_use; 9048 break; 9049 case TCP_RACK_TLP_INC_VAR: 9050 /* Does TLP include rtt variance in t-o */ 9051 optval = rack->r_ctl.rc_prr_inc_var; 9052 break; 9053 case TCP_RACK_IDLE_REDUCE_HIGH: 9054 optval = rack->r_idle_reduce_largest; 9055 break; 9056 case TCP_RACK_MIN_PACE: 9057 optval = rack->r_enforce_min_pace; 9058 break; 9059 case TCP_RACK_MIN_PACE_SEG: 9060 optval = rack->r_min_pace_seg_thresh; 9061 break; 9062 case TCP_BBR_RACK_RTT_USE: 9063 optval = rack->r_ctl.rc_rate_sample_method; 9064 break; 9065 case TCP_DELACK: 9066 optval = tp->t_delayed_ack; 9067 break; 9068 case TCP_DATA_AFTER_CLOSE: 9069 optval = rack->rc_allow_data_af_clo; 9070 break; 9071 default: 9072 return (tcp_default_ctloutput(so, sopt, inp, tp)); 9073 break; 9074 } 9075 INP_WUNLOCK(inp); 9076 error = sooptcopyout(sopt, &optval, sizeof optval); 9077 return (error); 9078 } 9079 9080 static int 9081 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) 9082 { 9083 int32_t error = EINVAL; 9084 struct tcp_rack *rack; 9085 9086 rack = (struct tcp_rack *)tp->t_fb_ptr; 9087 if (rack == NULL) { 9088 /* Huh? */ 9089 goto out; 9090 } 9091 if (sopt->sopt_dir == SOPT_SET) { 9092 return (rack_set_sockopt(so, sopt, inp, tp, rack)); 9093 } else if (sopt->sopt_dir == SOPT_GET) { 9094 return (rack_get_sockopt(so, sopt, inp, tp, rack)); 9095 } 9096 out: 9097 INP_WUNLOCK(inp); 9098 return (error); 9099 } 9100 9101 9102 struct tcp_function_block __tcp_rack = { 9103 .tfb_tcp_block_name = __XSTRING(STACKNAME), 9104 .tfb_tcp_output = rack_output, 9105 .tfb_tcp_do_segment = rack_do_segment, 9106 .tfb_tcp_ctloutput = rack_ctloutput, 9107 .tfb_tcp_fb_init = rack_init, 9108 .tfb_tcp_fb_fini = rack_fini, 9109 .tfb_tcp_timer_stop_all = rack_stopall, 9110 .tfb_tcp_timer_activate = rack_timer_activate, 9111 .tfb_tcp_timer_active = rack_timer_active, 9112 .tfb_tcp_timer_stop = rack_timer_stop, 9113 .tfb_tcp_rexmit_tmr = rack_remxt_tmr, 9114 .tfb_tcp_handoff_ok = rack_handoff_ok 9115 }; 9116 9117 static const char *rack_stack_names[] = { 9118 __XSTRING(STACKNAME), 9119 #ifdef STACKALIAS 9120 __XSTRING(STACKALIAS), 9121 #endif 9122 }; 9123 9124 static int 9125 rack_ctor(void *mem, int32_t size, void *arg, int32_t how) 9126 { 9127 memset(mem, 0, size); 9128 return (0); 9129 } 9130 9131 static void 9132 rack_dtor(void *mem, int32_t size, void *arg) 9133 { 9134 9135 } 9136 9137 static bool rack_mod_inited = false; 9138 9139 static int 9140 tcp_addrack(module_t mod, int32_t type, void *data) 9141 { 9142 int32_t err = 0; 9143 int num_stacks; 9144 9145 switch (type) { 9146 case MOD_LOAD: 9147 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 9148 sizeof(struct rack_sendmap), 9149 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); 9150 9151 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 9152 sizeof(struct tcp_rack), 9153 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 9154 9155 sysctl_ctx_init(&rack_sysctl_ctx); 9156 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, 9157 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 9158 OID_AUTO, 9159 __XSTRING(STACKNAME), 9160 CTLFLAG_RW, 0, 9161 ""); 9162 if (rack_sysctl_root == NULL) { 9163 printf("Failed to add sysctl node\n"); 9164 err = EFAULT; 9165 goto free_uma; 9166 } 9167 rack_init_sysctls(); 9168 num_stacks = nitems(rack_stack_names); 9169 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, 9170 rack_stack_names, &num_stacks); 9171 if (err) { 9172 printf("Failed to register %s stack name for " 9173 "%s module\n", rack_stack_names[num_stacks], 9174 __XSTRING(MODNAME)); 9175 sysctl_ctx_free(&rack_sysctl_ctx); 9176 free_uma: 9177 uma_zdestroy(rack_zone); 9178 uma_zdestroy(rack_pcb_zone); 9179 rack_counter_destroy(); 9180 printf("Failed to register rack module -- err:%d\n", err); 9181 return (err); 9182 } 9183 rack_mod_inited = true; 9184 break; 9185 case MOD_QUIESCE: 9186 err = deregister_tcp_functions(&__tcp_rack, true, false); 9187 break; 9188 case MOD_UNLOAD: 9189 err = deregister_tcp_functions(&__tcp_rack, false, true); 9190 if (err == EBUSY) 9191 break; 9192 if (rack_mod_inited) { 9193 uma_zdestroy(rack_zone); 9194 uma_zdestroy(rack_pcb_zone); 9195 sysctl_ctx_free(&rack_sysctl_ctx); 9196 rack_counter_destroy(); 9197 rack_mod_inited = false; 9198 } 9199 err = 0; 9200 break; 9201 default: 9202 return (EOPNOTSUPP); 9203 } 9204 return (err); 9205 } 9206 9207 static moduledata_t tcp_rack = { 9208 .name = __XSTRING(MODNAME), 9209 .evhand = tcp_addrack, 9210 .priv = 0 9211 }; 9212 9213 MODULE_VERSION(MODNAME, 1); 9214 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 9215